How Mistral Small 4 Unifies Reasoning, Vision, and Agentic Coding

Introduction

In modern AI deployments, engineers often face the compromise of either a large (and expensive), computationally intensive model for deep analytic problem-solving or a small (and inexpensive) rapid (and reflexive) model for response generation. In turn, these two models create a fragmentation of application, resulting in excessive costs associated with switching from tool to tool for each particular task. Inference costs mount up due to both of the above factors and backend orchestration becomes more difficult.

To address this problem of fragmentation, Mistral Small 4 focuses on optimizing enterprise-class performance by merging many unique analytic and generative capabilities into one cohesive engine. New developments and analyst reports show that no longer needing to route requests through independent models for the tasks of instruction, reasoning and visuals will allow enterprises the ability to support their own operational overheads while delivering frontier-class levels of intelligence through a single, optimized engine.

What is Mistral Small 4?

Mistral Small 4 is a powerful, unifying hybrid language model that combines the strengths of three formerly separate model families: Instruct, Magistral (for reasoning), and Devstral (for agentic coding). The new model is designed to be a versatile, all-around enterprise solution to remove the barriers associated with managing separate checkpoints.

Key Features of Mistral Small 4

• Unified Model Intelligence: The model can seamlessly combine instruction following, step-by-step reasoning, and agentic coding skills into one single engine, thus eliminating the need to switch between different models for different tasks.
• Reasoning on Demand: The programmable parameter reasoning_effort allows systems to toggle seamlessly between fast and low-latency responses (functioning like Mistral Small 3.2) and deep analytical step-by-step reasoning (functioning like Magistral models) in the very same model instance.
• Native Multimodality: Unlike its text-centric Small-family predecessors, Small 4 is built out-of-the-box to process text and image input simultaneously, thus allowing for complex visual document parsing and codebase exploration without needing to use any additional vision model.
• Frontier-Scale Context: The model includes a massive 256k context window, thus allowing for codebase ingestion and textual filings in a single inference turn. This is similar to what we see in the frontier Mistral Large 3.

Use Cases of Mistral Small 4

• Dynamic High-Density Multimodal Codebase Auditing: By utilizing the 256k context window and multimodal nature of the model, it allows for the simultaneous ingestion of massive application architectures and their corresponding visual user interfaces. This allows a developer to deploy this one engine to perform reflex-grade perception to identify visual elements in the UI and then immediately utilize the reasoning_effort parameter to perform in-depth, step-by-step logic to debug visual bugs through massive backend code.
• Cost-Optimized High-Throughput Legal-Visual Discovery: By being optimized to maximize accuracy with significantly fewer output characters, the model can now process massive amounts of scanned evidence and text-based filings with a significant reduction in token usage compared to larger open-weight models. This reduces the cost of ownership in data-intensive legal discovery processes.
• Unified Multimodal Agentic Supply Chain Surveillance: The model functions as an autonomous agent to monitor visual inventory feeds in real-time while concurrently analyzing text-based logistics logs in a single inference step. Its high sparsity enables it to utilize merely a small portion of its potential to achieve top-notch throughput in visual discrepancy identification in real-time.
• Low-Latency Visual Grounding for Interactive Desktop Agents: The unified model enables high-performance desktop automation agents to process high-definition visual inputs while executing intricate terminal commands. By maintaining benchmark-level performance while utilizing merely a small portion of its parameters, it ensures industry-leading latency in desktop automation while reasoning to solve intricate UI challenges.

How Does Mistral Small 4 Work?

Mistral Small 4 uses an innovative hardware-aware Mixture of Experts (MoE) sparse form of intelligence. The system has a massive capacity of 119 billion parameters, with its intelligence spread across 128 experts. Nevertheless, the system has been engineered with extreme sparsity, as for each input token, a sophisticated system only uses the top 4 experts out of the total number of experts. This means that, although the system has access to the vast knowledge base of a 119B model, it has an inference profile similar to a much smaller model, as only 6 billion parameters are used per token (8 billion including embedding and output layers).

The architectural design is further refined with Reinforcement Learning (RL) hybrid training on trillions of text and image tokens. It is specifically tuned to optimize performance per token, maximizing accuracy on benchmarks while forcing minimization of output length. To ensure enterprise scaling, the inference stack was co-developed with NVIDIA to ensure day 0 support in NVIDIA NIM, vLLM, and SGLang. The optimized serving stack maximizes hardware utilization and throughput with disaggregated GPU hardware while executing dynamic reasoning_effort toggles without any computational bottlenecks.

Performance Evaluation with Other Models

Mistral Small 4 has created a new benchmark for efficiency, with a fundamental improvement over heavyweights in the 80B-120B range of parameters with respect to accuracy per token. 

source - https://mistral.ai/news/mistral-small-4

On the AA LCR (Artificial Analysis Live Code Reasoning) benchmark, as emphasized in the evaluation metrics of Score vs. Output Length, Mistral Small 4 has a competitive score of 0.72 with only 1.6K characters of output text. This is a fundamental difference from the Qwen models, which require 3.5x-4x more output, i.e., 5.8K-6.1K characters, to attain comparable performance. This drastic reduction in verbosity also means a whopping 40% improvement in completion time as well as a 3x improvement in throughput over Mistral Small 3, which has a direct bearing on inference cost mitigation.

source - https://mistral.ai/news/mistral-small-4

In addition to this, the model shows unprecedented sparsity during the evaluation of Live Code Bench. The Mistral Small 4 achieves or outperforms the frontier-level GPT-OSS 120B on intricate coding and agentic problems while producing 20% less output volume. This achievement of parity with the dense 120B model while using only 6 billion parameters per token shows the superiority of the 128-expert MoE architecture. This means that deep reasoning and code generation can be done with significantly shorter, highly accurate tokens, resulting in substantial savings on the cost of intervention during lengthy, hallucinatory outputs of the previous generation.

How to Access and Use Mistral Small 4?

Mistral Small 4 is released under the liberal Apache 2.0 license, which allows free and unimpeded open-source and commercial use. The model is accessible by direct access to the model weights via the Hugging Face repository, which contains configurations to support optimized serving frameworks like vLLM and SGLang. The model is also available natively via the Mistral API, AI Studio, and containerized NVIDIA NIM (testable via an online demo at build.nvidia.com). The recommended solution for self-hosting on local infrastructure is disaggregated inference across 16x NVIDIA H200, with minimum hardware configurations starting at 4x H100, 2x H200, or 1x B200.

Limitations 

While the model has excellent efficiency, its total parameter count of 119B makes infrastructure a major challenge; the model will not run on common laptops or consumer-grade GPUs. In addition, there might be performance trade-offs with extremely quantized checkpoints, especially with extremely long context scenarios, compared to full-precision versions, like NVFP4. While not unique to generative models, human validation will always be required for critical decision-making. 

The Next Efficiency Frontier?

The move toward a hybrid framework will clearly open some intriguing possibilities for the future of scaling the enterprise. Will this single-engine strategy finally allow for the elimination of the brittle middleware that is currently required to shuttle between the disconnected vision and coding paradigms? By using the dynamic control of depth for analysis, there is a tremendous opportunity to develop self-optimizing layers of orchestration that can dynamically apply lower intensity to routine queries while still holding higher intensity for complex logical anomalies. This is clearly a strategy that will revolutionize the management of the inference budget without sacrificing the overall fidelity of the system.

Additionally, the development of hardware-optimized sparse architectures points to a future in which disaggregated configurations become the standard in high-performance private clusters. What if a model can deliver frontier-level precision with only a small percentage of its potential active? Will we soon witness highly responsive multimodal agents that are always online and operate at near-zero latency in secure corporate environments? Will this be the moment when open-weight efficiency finally supplants proprietary APIs in high-stakes data-sensitive automation? The unification of massive context and extreme character efficiency points to a future in which thinking is no longer simply a static attribute, but rather a dynamic and flexible resource.

Conclusion

In its attempt to bring reasoning, vision, and agentic coding under a single, very optimized framework, it has successfully addressed the fragmentation of deployment problem. To those working on scalable AI pipelines, a model that can reach 120B-level intelligence, using only 6B parameters per token, is not a trivial software update, but a re-architecture of how we compute the economics, latency, and business predictability of high-level AI work.

Sources:
Blog: https://mistral.ai/news/mistral-small-4
Model Weight: https://huggingface.co/collections/mistralai/mistral-small-4
Document: https://legal.cms.mistral.ai/assets/d0b7b04d-dcb5-412d-bb45-c63b1475b805

Disclaimer - This article is intended purely for informational purposes. It is not sponsored or endorsed by any company or organization, nor does it serve as an advertisement or promotion for any product or service. All information presented is based on publicly available resources and is subject to change. Readers are encouraged to conduct their own research and due diligence.

NVIDIA Nemotron 3 Super: Redefining Multi-Agent Enterprise AI

Introduction

The world of sparse intelligent and enterprise modeling has just experienced a revolution. As the world of intelligent and autonomous system pipelines continues to stretch the boundaries of goal-oriented logical planning and decision-making, the need for logical consistency within an unbound context without the burden of compute costs has never been greater. Traditionally, the standard dense network infrastructure has had trouble handling the immense burden of the so-called 'thinking tax,' where the costs of computation increase exponentially with the increase in the complexity of the task at hand. However, the latest innovation in the world of sparse network design is a fundamental shift in this concept. 

Through the utilization of a hybrid state space and attention mechanism with a dimension-reduced sparse expert network, a new frontier in the world of logical planning and function integration is being realized. This is particularly true in the realm of integrating complex logical actions and advanced function integration within a real-time capacity. Whether the need is to integrate monolithic software migrations, server clusters with bandwidth constraints, or zero-latency evaluation environments, Nemotron 3 Super is the embodiment of this revolution in the world of logical reasoning and decision-making.

What is Nemotron 3 Super?

Nemotron 3 Super is a very efficient 120 billion parameter open weight model that uses only 12 billion parameters per forward pass. The model is created by NVIDIA and is the latest in the line of sparse reasoning engines. The model is specifically designed to function as the cognitive core of complex, multi-agent enterprise applications. The model is competing very fiercely with similar-class models like GPT-OSS-120B and Qwen3.5-122B and even much larger trillion-parameter models.

Key Features of Nemotron 3 Super

• Native NVFP4 Pre-training: The model has been natively pre-trained across 25 trillion tokens in 4-bit floating-point numbers using the Blackwell architecture. This removes the post-hoc quantization sensitivity and increases the inference speed by up to 4x compared to FP8.Lynx + LLM.
• Mathematical Integrity Pipeline: The model uses a text-based browser called Lynx for rendering HTML documents before the text is processed. Then, a teacher model normalizes all notation into strict LaTeX. This prevents formatting noise from corrupting the data.
• PivotRL for Turn-Level Optimization: This is a turn-level reinforcement learning approach that heavily relies on the Supervised Fine-Tuning (SFT) trajectories. However, it has a high focus on ambiguous pivot points. This creates a policy that is excellent at handling ambiguity without out-of-distribution issues.
• Group Relative Length Control: This is an integrated length penalty feature used in RLHF. It is intended for controlling verbosity bias by ensuring that the response is always accurate and brief. This feature is useful for reducing the amount of tokens used by the enterprise.
• Shared-Weight MTP Heads: This feature harmonizes all the prediction heads of the model, thus eliminating training-inference divergence, which is common in standard Multi-Token Prediction (MTP) models, to ensure elongated drafts during speculative decoding.

Use Cases for Nemotron 3 Super

• Real-Time Industrial System Reliability Engineering (SRE): The model is able to call upon 22 specialist experts for every token input. This allows it to process disparate telemetry streams in a single forward pass. This is 7.5 times faster than equivalent models and thus provides the necessary speed for real-time intervention in smart factories.
• Monolithic Codebase Deep Reconstruction: The model is able to hold the global intent of entire legacy codebases in its 1M token context window. This allows autonomous systems to rewrite low-level functions without losing the architectural thread.
• Distributed Strategic Modeling in Bandwidth-Constrained Clusters: The dimension-reduced latent routing slashes the payload required for all-to-all communication that is normally necessary in standard MoE models. This allows robust 120B-class reasoning to occur in legacy datacenter tiers or even non-specialized compute nodes without suffering a throughput collapse.
• Native 4-bit Private Enterprise Intelligence: The model is able to run sensitive corporate workflows such as legal discovery on a single GB200 workstation. This is frontier-level reasoning capability without the zero-valued weight gradients that normally plague compressed high-parameter models.
• High-Stakes Autonomic IT Automation: Leveraging PivotRL training capabilities, the model efficiently completes routine networking tasks at maximum speed in low-effort mode and adapts to high-accuracy reasoning in high-stakes decision points such as uncertain security threats and novel attack methods.

How Does Nemotron 3 Super Work?

The inner workings of Nemotron 3 Super's machinery are based upon a Latent Mixture-of-Experts architecture. This is in contrast to traditional routing methods, as the architecture projects tokens into a 1024-dimensional space for expert calculation. This allows for a reduced payload in the communication by a factor equal to the dimensionality of the model compared to its latent space compression. This allows the model to activate 4 times as many experts for the same cost as before, resulting in a much higher accuracy per byte generated.

source - https://research.nvidia.com/labs/nemotron/files/NVIDIA-Nemotron-3-Super-Technical-Report.pdf

The sparse routing is based upon a Hybrid Interleaved Pattern within an 88-layer stack. 

source - https://research.nvidia.com/labs/nemotron/files/NVIDIA-Nemotron-3-Super-Technical-Report.pdf

The sparse LatentMoE layers are interleaved with linear-time Mamba-2 layers to alleviate the quadratic cost of massive KV caches. Additionally, the global Transformer layers are used as a form of  logical anchor for the architecture. The alignment is also guided by a massive 235B parameter Generative Reward Model (GenRM), which is used to rank reasoning traces with the level of precision typically reserved for closed-source models.

Performance Evaluation with Other Models

In the case of throughput and long-horizon software engineering benchmarks, Nemotron 3 Super exhibits overwhelming dominance. It obtains an unprecedented result of 60.5% on the SWE-bench. This is a significant improvement compared to the 38.8% obtained by the Nano variant. This justifies its capability for deep codebase reconstruction and complex terminal usage. In the case of mathematical reasoning and exploratory writing, the performance of the model is also higher than that of other variants and larger architectures. On the challenging HMMT Feb 25 (Math) benchmark, the model obtains an unprecedented result of 94.7% accuracy. This is a significant improvement compared to the 90.0% accuracy obtained by the GPT-OSS-120B variant.

source - https://research.nvidia.com/labs/nemotron/files/NVIDIA-Nemotron-3-Super-Technical-Report.pdf

Moreover, in the case of generation-intensive usage scenarios (8k input and 64k output), the model obtains 7.5 times higher throughput than the Qwen3.5-122B variant and 2.2 times higher throughput than the GPT-OSS-120B variant. This justifies its dominance in the case of massive output usage scenarios without the potential for enterprise compute bottlenecks.

source - https://research.nvidia.com/labs/nemotron/files/NVIDIA-Nemotron-3-Super-Technical-Report.pdf

Further, the Shared Weight MTP Heads have the highest average acceptance length of 3.45 tokens, surpassing DeepSeek-R1's 2.70 tokens on SPEED-Bench, even for speculative decoding. This leads to a radical acceleration of token generation, especially during highly complex reasoning traces.

How to Access and Use Nemotron 3 Super?

The model is openly accessible in the form of the model weights in BF16, FP8, and NVFP4 data formats through the NVIDIA Open License on the Hugging Face website. The endpoint is immediately accessible through build.nvidia.com, OpenRouter, and Perplexity. The model is provided as an NVIDIA NIM microservice, thus ensuring easy integration through orchestration tools like vLLM, SGLang, and TRT-LLM. For the robust local deployment of the model, the usage of the multi-agent frameworks, and the deep open-source customization of the system, the primary source of information should be the recipes provided in the NVIDIA Nemotron Developer GitHub repository. The repository contains the methodologies required to compile and run the system locally in a secure environment.

Limitations

 Even with its high degree of performance, it is still very much dependent on hardware; i.e., the NVIDIA Blackwell platform must be used exclusively in order to achieve the 4X speedup found in NVFP4. In addition, while quantization sensitivity has been identified during model training (i.e., 7% of zero-weight gradient weights require special recipes to ensure that the system remains stable), the decoupled DRAM reads of Mamba State Cache data resulted in an additional 37%-40% total overhead (i.e., verbosity spike), which will be addressed through stochastic rounding.

Conclusion

Nemotron 3 Super is not simply an increase in the number of parameters; it represents a master class in the development of hardware-aware software. By reducing the routing payload to a latent space and training natively in 4-bit precision, it successfully addresses the interconnect bandwidth issues that most MoE models of this size experience. For technical teams who are working with limited IT budgets but need to leverage cutting-edge reasoning to operate complex monolithic code bases or SRE telemetry at high throughput, this model eliminates the need to compromise accuracy for latency and indicates that future advances in the scalability of AI will be driven by activation efficiency rather than parameter count alone.

Sources:
Blog: https://blogs.nvidia.com/blog/nemotron-3-super-agentic-ai/
Model Weight: https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-FP8
Technical document: https://research.nvidia.com/labs/nemotron/files/NVIDIA-Nemotron-3-Super-Technical-Report.pdf

Disclaimer - This article is intended purely for informational purposes. It is not sponsored or endorsed by any company or organization, nor does it serve as an advertisement or promotion for any product or service. All information presented is based on publicly available resources and is subject to change. Readers are encouraged to conduct their own research and due diligence.

Gemini Embedding 2: Direct Multimodal Search Without Text Conversion

Introduction

The future of how semantic search and intelligent data retrieval work has shifted from treating each type of media as independent and separate 'silos', For those architecting a new data system, this evolution has already transitioned to a more fluid approach. There are different pillars that define this evolution: the integration of a naturally unified process to ingest sensory data;  the mapping of disparate data streams into a cohesive multi-dimensional vector based on informed similarities;  the use of dynamic vector scaling in order to ensure that the storage cost of the vector is balanced against the precision of the retrieval; and guidance on how query algorithms should interpret the user's search intent depending on the type of statistical query model being employed.

The adoption of Gemini Embedding 2 is driven primarily by its ability to collapse technical debt. By removing the traditional 'transcribe then index' bottleneck associated with video and audio content, it significantly speeds up the time to insight for both types of content while ensuring that the semantic subtleties which can often be lost during the transcription process are preserved. This also creates a single, high-performing system through which video, audio, and text-based information can be combined in a seamless manner.

What is Gemini Embedding 2?

Gemini Embedding 2 is Google's first ever multimodal embedding model and is intended to function as the fundamental cognitive layer for Retrieval-Augmented Generation models of a higher order as well as massive data management. By mathematically uniting completely different data formats within a single common geometric space, Gemini Embedding 2 enables complicated relationships that are cross-modal to be innately comprehended and queried without the need for traditional text-centric translation constraints.

source - https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-embedding-2/

Key Features of Gemini Embedding 2

• Massive Context Window Expansion: The model now has an input limit of 8,192 tokens. This is a huge jump from the 2,048 token input limit of its previous incarnation. The result is that it can now handle larger chunks of code, document snippets, and other contextual data within a single input operation without requiring any kind of chunking operation.
• Interleaved Input Understanding: Legacy models require that visual data be split from text data prior to input. Gemini Embedding 2 can handle interleaved data within a single API call. In other words, it is able to successfully map sequential and relational data between text paragraphs and images within a single input operation.
• Advanced Document and Media Handling: The Gemini Embedding 2 has native Document OCR capabilities that allow it to read document text directly from PDFs. Additionally, it has audio track extraction capabilities that allow it to extract audio from videos to interleave it with visual data.
• Expansive Multilingual Support: For global enterprises that require multilingual knowledge retrieval, Gemini Embedding 2 has native multilingual support for more than 100 languages. This is a huge advantage for those who require a rapid solution for multilingual data.

Use Cases of Gemini Embedding 2

• Streamlined Multimedia Audit and Discovery: Media firms, legal discovery teams, and archivists can search vast and untapped digital media archives to find specific video scenes or audio segments using just a simple query based on the description of the scene or the reference audio bite itself.
• Intelligent Technical Document Retrieval (Visual RAG): Technical teams in the fields of engineering, medicine, and law can develop accurate RAG systems that retrieve critical information embedded within complex PDF layouts. This way, experts can instantly retrieve architectural diagrams, medical charts, and financial tables that might be missed by text parsers.
• Context-Aware Sentiment Monitoring: Brand management and marketing teams can accurately measure the sentiment of the public on social media by analyzing the content of social media posts where the meaning of the post is heavily influenced by the interaction of the media types. For example, teams can successfully identify the sentiment of the post where the meaning of the post changes completely due to the presence of an image that is sarcastic and the text caption is positive.
• Cost-Optimized Global Search Engines: E-commerce sites and multinational companies can create blazingly fast and highly relevant search experiences for products and content in global markets, all while minimizing storage and compute costs on the vector database.
• Specialized Code Knowledge Bases: Software development companies can create internal developer portals where junior developers can ask natural language questions and get instant access to the exact corresponding proprietary code blocks or system architecture schemas.

How Does Gemini Embedding 2 Work?

From a software architecture point of view, the workflow of the Gemini Embedding 2 system is significantly different from the standard sequential workflow. The most significant difference in the software architecture of the system is the ingestion of raw audio data. Unlike the standard workflow of ingesting raw data through the ASR engine and producing intermediate text transcripts, the system ingests raw audio data. This way, the semantic nuances of the raw data are not lost during the ingestion process.

The mathematical core of the system is based on Matryoshka Representation Learning (MRL). MRL is a training method that nests the information. It optimizes the loss function on multiple levels simultaneously. Due to the presence of the MRL method, the developers are not required to use the standard 3072-dimension vector. They are allowed to truncate the vector to a lower dimension size, such as 1536 dimensions or 768 dimensions.

However, there is a critical architectural caveat: embedding incompatibility. Due to the fundamental difference in the geometric mapping of the unified space from the text-only architecture of the previous gemini-embedding-001, the two embedding spaces are mutually incompatible. To upgrade from the previous embedding to the new Gemini Embedding 2, all previous historical data must be re-embedded; it is not possible to transform the previous vectors to the new vectors.

Performance Evaluation with Other Models

When tested against some of the best-performing models currently used in the industry, Gemini Embedding 2 sets a whole new standard for multi-modal depth, particularly in tasks that involve cross-modal reasoning between text, image, and video data. Perhaps one of the greatest achievements in its testing is its MRL performance stability. As tested through standardized evaluations such as the Massive Text Embedding Benchmark (MTEB), this model shows that truncation does not necessarily ruin efficacy. For example, if its MRL dimension is reduced from a hefty 2048 dimensions (scoring 68.16) to a much smaller 768 dimensions (scoring 67.99), then such a reduction in quality is utterly negligible. This shows that systems can save massive amounts of compute and storage without compromising retrieval accuracy.

source - https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-embedding-2/

As a second evaluation vector, although certainly not less important, is its formidable speech capabilities. By bypassing traditional ASR systems, Gemini Embedding 2 introduces unique acoustic reasoning capabilities that show a statistically significant improvement over legacy foundation models. Such capabilities allow for acoustic semantics that are simply not possible to perceive through text or retrofitted multi-modal systems.

Competitive Benchmarking

To effectively contextualize the above benchmark results in the embedding world, you will need a strong sense of how Gemini Embedding 2 compares with heavyweights like Amazon Nova 2 and Voyage Multimodal 3.5. While Voyage Multimodal 3.5 has the strongest capacity for RAG with its massive 32K token context window, allowing for successful RAG on book length documents, its acoustic capability does not begin to touch the capability of the Gemini Ecosystem. While Amazon Nova 2 offers a strong five modality space, such as text, images, audio, etc., with highly aggressive truncation options as low as 256 dimensions, its media input restriction of 30 seconds results in a fragmented, chunked ingestion methodology. In contrast, Gemini Embedding 2 finds a middle ground with its focus on semantic continuity, offering an 8K token context window with the best temporal fidelity possible, supporting 120 seconds of video and 80 seconds of native audio in one unchunked request.

Gemini Embedding 2 reconfirms its position as first choice for high-latency reasoning and cross-modal semantic integrity. Skipping over the entire ASR pipeline, Gemini Embedding 2 is able to tap into the 'soul' of the audio data that is simply never available in the text-based pipelines of the competitors. This ensures that regardless of the query being performed on a two-minute scene or complex data sheet, the model has a cohesive semantic map that is simply never available to the 30-second-limited competitors. This is a choice for the architect: the sheer volume of Voyage, the storage efficiency of Nova, or the semantic integrity of Gemini.

How to Access and Use Gemini Embedding 2?

As of March 10. 2026, Gemini Embedding 2 is accessible for business use through the Gemini API as well as Vertex AI. The Gemini Embedding 2 model is accessible for use within a variety of major ecosystem integrations that are presently available. Currently, infrastructure access is strictly limited to a Standard PayGo consumption model; high-volume business features like Provisioned Throughput as well as Batch Prediction are not yet accessible during this time.

Limitations and Future Work

While the 'as is' preview release of the architecture has many strengths, it also has strict input limits per request. This includes up to 6 images, 120 seconds of video (or 80 seconds if the video contains audio), 80 seconds of audio, and up to 6 pages of PDF. Additionally, the architecture is geographically restricted to the 'us-central1' region. However, the architecture itself is meant to be the foundation of the future of the evolution of context engineering. Therefore, it is expected that the limits of the architecture will increase as the architecture itself evolves to handle more multimodal RAG and data management needs.

Conclusion

For teams working with substantial scale, this ability to truncate dimensions while still maintaining a score close to a MTEB score means that you can literally cut your vector database hosting costs in half overnight. Although the heavy upfront effort necessary for migrating and re-embedding existing databases will be necessary, the ability to perform unified visual, acoustic, and text-based searches in one action will make it essential for serious data infrastructures.

Sources:

Blog: https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-embedding-2/

Vertex API: https://docs.cloud.google.com/vertex-ai/generative-ai/docs/models/gemini/embedding-2

Gemini API document: https://ai.google.dev/gemini-api/docs/embeddings

Disclaimer - This article is intended purely for informational purposes. It is not sponsored or endorsed by any company or organization, nor does it serve as an advertisement or promotion for any product or service. All information presented is based on publicly available resources and is subject to change. Readers are encouraged to conduct their own research and due diligence.

Phi-4-Reasoning-Vision-15B: Microsoft's Open-Weight Multimodal AI

Introduction

As artificial intelligence continues its journey from text-centric interfaces into the visually complex world, new models are emerging that bridge the gap between simple perception and deep logic. Until recently, this has necessitated massive computational overheads, which has restricted innovations to enterprise server farms. However, this is no longer an issue in modern compact multimodal models, which get their power from a sequence of innovations rather than massive parameterization. First, combined analytical processing enables dynamic routing that adjusts computational depth in real time according to complexity requirements. Second, meticulous dataset curation ensures that these models are trained on pristine data quality, which is paramount over quantity. Third, structural innovations have provided the necessary bridges between visual and text inputs without compromising detail.

Why develop a compact powerhouse like Phi-4-Reasoning-Vision-15B today? The tech world is running into a wall in terms of finances and compute capability with monolithic models. We need tools that move the Pareto-frontier in terms of efficiency, providing high-fidelity actionable intelligence without needing astronomical compute times or token generations to reach our goals.

What is Phi-4-Reasoning-Vision-15B?

Phi-4-Reasoning-Vision-15B is a unique small language model that has been optimized for both text and visual reasoning. As a cognitive engine with the capability of interpreting complex images, locating tiny parts of these images, and making logical deductions through multiple steps, this architected model also has one of the smallest operational footprints in the industry.

Key Features of Phi-4-Reasoning-Vision-15B

• Selective Task-Aware Reasoning: It has the native ability to switch between two very disparate modes of operation. It has the capacity to employ a chain of thought process, initiated by think tags, to solve problems in a multi-step manner, and a direct response process, initiated by nothink tags, to solve problems in a low-latency manner.
• High-Resolution GUI Grounding: It is natively optimized to solve Computer Using Agent (CUA) problems, in which it has the capacity to interpret the densely populated digital world. It has the capacity to precisely identify interactive objects like menus, icons, and buttons, and translate them into exact coordinate-based actions.
• Scientific and Mathematical Visual Reasoning: While other systems are limited to the recognition of simple images, this model is capable of solving complex mathematical problems presented in the form of diagrams and accurately interpreting dense mathematical data presented in the form of complex and convoluted charts and tables.
• Sequential Image Interpretation: While other systems are limited to the interpretation of a single image in a vacuum, this groundbreaking feature has the capacity to analyze the changes between a series of images, and interpret the manner in which a given situation or object has evolved.

Use Cases of Phi-4-Reasoning-Vision-15B

• Automated Troubleshooting in High-Density GUI Items: The model serves as an agent in very complicated legacy software structures (e.g., multi-layer trading workstations and financial dashboards). The model uses visual information to move through a series of complex displays, making precise motions based on coordinates to fix some of the state problems that will not be able to be fixed using standard back-end APIs.
• Real-Time Diagnostics of Physical Infrastructure Maintenance: Predictive maintenance can be achieved by analyzing the changes of an industrial component's visual state over an extended period of time (across several consecutive images) and by understanding the succession of mechanical failure-based logical progression rather than treating each image separately.
• High Quality Document Intelligence: The model has the ability to effectively process high-quality documents that use many pages and have a high-resolution image quality (e.g., a medical record, the various annotations associated with each X-ray, and civil engineering-related documents). The model is able to preserve detailed information in order to create a reliable visual audit report of the symbols used within each document for subsequent validation (e.g., digitization of diagrams).
• Optimally Reducing Latency in Hybrid Mobile Navigation: In both mobile and IoT environments, the model is able to recognize and use data to quickly locate application icons, while also using previously accumulated reasoning when executing a user command that requires complex visual/spatial reasoning.

How Does Phi-4-Reasoning-Vision-15B Work? 

At a high level, the architecture of Phi-4-Reasoning-Visual-15B is based on a highly efficient Mid-Fusion Architecture. The way in which this architecture works is to use a pre-trained SigLIP-2 vision encoder to convert the raw input image into a series of visual tokens. Then, after the vision tokens have been generated, a cross-modality projector will project the vision tokens directly into the embedding space of the pre-trained Phi-4-Reasoning language backbone. This method is far more computationally efficient than using an early fusion method, effectively allowing for the use of two foundational models, both of which have been trained on trillions of tokens, without having to construct them from the ground up. 

source - https://www.microsoft.com/en-us/research/wp-content/uploads/2026/03/Phi-4-reasoning-vision-15B-Tech-Report.pdf

Another important structural innovation is the inclusion of the SigLIP-2 NaFlex dynamic resolution variant. This mechanism is designed to accommodate variable visual inputs. It is capable of producing up to 3,600 visual tokens per image, equivalent to the native resolution of HD 720p. This dynamic scaling is important in ensuring that the model is able to grasp even the most microscopic details in dense screenshots or schematics that traditional encoders would normally blur or ignore. The training process is also highly specialized, involving a targeted Hybrid Training Mixture. The model only consumes a mere 200 billion tokens of multimodal data, a small fraction of the trillion-token diets of rival models such as Qwen 3 VL or Gemma 3.  A very important innovation is the imposition of a very strict hallucination mitigation protocol. Unlike earlier models that are prone to improv-style guessing, the current model is explicitly trained to fail to produce an answer when factual certainty is below a certain threshold.

Performance Evaluation with Other Models

The performance of the model’s interface grounding capacity was extensively tested with the ScreenSpot-v2 benchmark, as elucidated in the performance reviews in tables below. For the particular domain, the Phi-4-Reasoning-Vision-15B model was able to obtain a remarkable performance of 88.2%. This is a tremendous evolutionary leap from its previous version, the Phi-4-mm-instruct, which was only able to obtain a dismal performance of 28.5%. The benchmark also evaluates the unprecedented capacity of the model to accurately pinpoint minute interactive elements on the screen, outperforming larger models from the same company in direct screen manipulation.

source - https://techcommunity.microsoft.com/blog/azure-ai-foundry-blog/introducing-phi-4-reasoning-vision-to-microsoft-foundry/4499154

For the complex mathematical logic problem, the performance of the model was assessed with the MathVista and MathVision benchmarks. The performance of the model was also superior for complex mathematical logic when compared to similarly fast open-weight models, thus validating the effectiveness of the synthetic data strategy for reasoning. The model was able to push the Pareto frontier of efficiency, thus demonstrating its high competitiveness with models that are ten times larger in terms of parameters and have a much larger compute time as well as token generation overhead.

source - https://techcommunity.microsoft.com/blog/azure-ai-foundry-blog/introducing-phi-4-reasoning-vision-to-microsoft-foundry/4499154

Apart from the above primary tests, the model has retained its robust competitiveness on all other broad vision language tests. In all additional tests, the internal switching logic of the model was found to be highly robust. It was observed that the model, on an average, performed better when it was allowed to be in its default mixed-reasoning state rather than being forced into either thinking or non-thinking modes, thus reiterating its position as an exceptionally balanced multimodal reasoning engine.

How to Access and Use Phi-4-Reasoning-Vision-15B

Deployment of this model can be done flexibly across several different platforms including Microsoft Foundry, Hugging Face and GitHub. The code is made available under the highly permissive MIT license. Users wishing to utilise managed infrastructure to deploy(such as Azure AI Foundry) will be able to deploy without needing to manage complex hardware. Those who wish to run locally may do so through either Hugging Face Transformers or vLLM frameworks, with the main source of information on how to do so being found chiefly through the official GitHub repository.

Limitations 

The model has a number of limitations despite having made an enormous amount of progress since it was first introduced. For example, the implicit boundary that determines the sub-optimal switching of modes between reasoning and responding is sometimes not accurate enough, and users have to manually override the model by using explicit tags in certain scenarios. In addition, the model has built-in weaknesses for following strict instructions. Specifically, it sometimes has trouble creating complex tables or specific bulleted items when compared to larger LLMs that are designed to follow instructions. The model is also limited in its overall ability to store data internally because of its compact design and can produce ‘hallucinated’ factual information concerning obscure facts or persons unless the model is being used in conjunction with a Retrieval-Augmented Generation (RAG) pipeline. 

Future Horizons: What’s Next for Compact Multimodal Engines?

Moving forward, we will increase the capabilities of compact reasoning engines. One possibility is to build on the Mixture-of-Experts (MoE) model as a core part of language architecture. By directing visual tokens to very particular expert pathways in the neural network, can we greatly increase the knowledge storage of the engine without adding VRAM at the edge? This would provide a way to overcome the factual limitations currently seen, but also continue to provide the zero-latency, local deployments needed for autonomous physical systems and disconnected networks.

Also, as the dynamic switch logic improves, sequential visual analysis may evolve into agentic (independent) and multi-step (several steps) behaviors. It may also be possible for this framework to not only identify problems in the interface of a logical system, but to automatically repair the logic and provide real-time updates/patches for complex legacy systems. If selective reinforcement learning could be applied to resolve idiosyncrasies in following instructions, will that enable an engine to manage visual and logical records on its own? The result will be to change this compact reasoning engine from a reactive analytic tool to an autonomous/self-repairing digital engine.

Conclusion

In promoting the application of Dynamic Resolution, High-Fidelity Data Curatorship and Selective Reasoning as opposed to only sheer parameter count, it provides a sustainable model that allows for the integration of profound analytical intelligence into local hardware, edge devices and legacy systems. As such, it demonstrates how efficiency and high levels of accuracy can coexist to provide an essential resource for users wanting to create strong visually based applications without the mass of overhead generated from traditional deep learning paradigms.

Sources:
Tech community Blog: https://techcommunity.microsoft.com/blog/azure-ai-foundry-blog/introducing-phi-4-reasoning-vision-to-microsoft-foundry/4499154
Research Blog: https://www.microsoft.com/en-us/research/blog/phi-4-reasoning-vision-and-the-lessons-of-training-a-multimodal-reasoning-model/
Tech document: https://www.microsoft.com/en-us/research/wp-content/uploads/2026/03/Phi-4-reasoning-vision-15B-Tech-Report.pdf
Model Card: https://huggingface.co/microsoft/Phi-4-reasoning-vision-15B
GitHub Repo: https://github.com/microsoft/Phi-4-reasoning-vision-15B


Disclaimer - This article is intended purely for informational purposes. It is not sponsored or endorsed by any company or organization, nor does it serve as an advertisement or promotion for any product or service. All information presented is based on publicly available resources and is subject to change. Readers are encouraged to conduct their own research and due diligence.

Qwen3.5: Scaling 17B Activation for Expert Visual Coding Logic

Introduction

Artificial Intelligence is now advancing towards self-regulating intelligent agents which will be able to work independently of human supervision and be able to carry out logical operations of multiple steps on their own. One key requirement for these new types of AI is that visual and semantic processing must be unified. Therefore, spatial information and textual information should be processed together as one continuous process in order for intelligent systems to be independent from each other. The new hybrid models allow for this kind of processing, enabling AI to achieve world-class cognitive performance levels, without requiring an inordinate amount of expense to operate high-density computational systems as was necessary in earlier generations.

This new AI technology has achieved cross-generational parity to the extent that it can now compete against trillion-parameter intelligence with only a fraction of the computational resources that were previously required to support dense-model intelligence. Because of its sparse and dual-modality design, extremely long context-length models may now be scaled directly, ending both the latency and infrastructure costs that typically accompany high-end AI capabilities such as advanced spatial reasoning abilities and automated code generation. This latest AI is referred to as Qwen3.5.

What is Qwen3.5?

Qwen3.5 is a strategic native vision-language foundation model designed to work as a holistic multimodal digital agent, not merely a tactical coding helper. It is developed using an early fusion training approach that processes trillions of diverse tokens in a single pass. This enables it to natively  see  and  think  at the same time, filling the gap between basic spatial perception and intricate logical computation.

Key Features of Qwen3.5

• Native Multimodal Fusion: In contrast to the previous versions that used separate encoding, Qwen3.5 uses early fusion training on trillions of multimodal tokens. This gives the model a baseline capability to perform expertly at Visual Coding—a capability that allows it to easily translate static UI sketches into functional and executable code or even reverse-engineer programmatic logic directly from recorded gameplay footage. It fundamentally grasps the causal connection between visual state transitions and software logic.
• Extreme Inference Efficiency: The Qwen3.5-397B-A17B flagship model has an enormous 397B total parameters but switches on only 17B parameters per forward pass. This is an unparalleled sparsity that gives it a decoding throughput of 19.0x faster than the >1T parameter Qwen3-Max-Base and 7.2x faster than the Qwen3-235B-A22B with a 256k context size.
• Massive Scalable RL Generalization: Moving away from the conventional scaled reinforcement learning approach that is designed to work easily in coding problems that can be readily verified, Qwen3.5 employs a disaggregated and asynchronous reinforcement learning approach. This allows the development of million-scale agent frameworks, which significantly increases its flexibility when deployed in real-world scenarios.
• Spatial Intelligence : The model has the capability to natively employ advanced pixel-level spatial relationship modeling. By doing so, it is able to counteract reasoning errors that normally take place as a result of perspective transformations in video or physical spaces.
• Superior Global Accessibility: In response to the requirement for superior global deployment, the linguistic ability has been significantly enhanced. The model is now capable of supporting 201 languages and dialects, which is a huge improvement over the 119 languages and dialects supported by Qwen3 and the 92 languages and dialects supported by Qwen2.5-Coder.

Use Cases of Qwen3.5

• Autonomous Logic Recovery from Legacy Dynamic Visual Systems: For projects involving the revival of outdated,  black box  legacy systems where the source code is either undocumented or completely lost, Qwen3.5 presents a paradigm shift. Based on the observation of operational videos or gameplay, the model uses its early fusion training to infer the logic structure by reverse-engineering the system. It deciphers the visual state transitions and expresses them in terms of the original causal programmatic logic, which can then be recovered solely through the observation of user interaction videos.
• Hyper-Scale Multi-Regional  Thinking:  Digital Workforce Organizations with the need for synchronized, worldwide digital forces can take advantage of the model’s million-scale agent frameworks. By delivering 19.0x the decoding throughput of bigger models at repository-scale context sizes, organizations can deploy millions of agents simultaneously. These agents can work in the default  thinking mode, performing structured reasoning on 262k+ token workflows in more than 200 dialects in real-time.
• Zero-Latency Multimodal Hardware-Optimized Edge Deployment: For infrastructure engineers building high-density clusters, Qwen3.5 is a game-changer. The model’s built-in FP8 pipeline and parallelism techniques provide a ~50% cut in activation memory. This enables the execution of repository-scale (1M+ token) visual coding tasks on much lighter hardware configurations, eliminating the Out-of-Memory (OOM) issues that come with traditional dense deployments.
• Automated Global Rebase and Visual-to-Logic Repository Maintenance: As a single multimodal project manager, the model can be used in conjunction with the Qwen Code CLI to manage enormous multi-language code repositories. With its 250k enhanced vocabulary and Efficient Hybrid Attention, the model can automate difficult repository rebases while performing visual checks on the integrity of the front-end UI in real-time, building without the latency issues of previous models.

How Does Qwen3.5 Work?

The main engine responsible for its speed and low latency is an Efficient Hybrid Architecture. This architecture replaces the usual attention mechanisms with a highly optimized combination of Gated Delta Networks for linear attention, Gated Attention, and a sparse Mixture-of-Experts  configuration. In particular, the hidden state configuration has a strict structure: Here is the hierarchy of the model's 'thinking' process.

15 Master Repetition Blocks, each containing:

• 3x Primary Logic Sub-blocks: Gated DeltaNet --> Mixture-of-Experts (MoE)
• 1x Contextual Integration Sub-block: Gated Attention --> Mixture-of-Experts (MoE)

In its working, it uses only 10 of the 512 experts available in the 397B parameter space for the routing mechanism per forward pass, limiting the active parameters to 17B. For handling the input data, it uses a Next-Generation Training Infrastructure that fully decouples the parallelism strategies for language and vision. This heterogeneous paradigm provides near-100% multimodal training efficiency relative to traditional text-only models. Moreover, it supports a 262K token context window natively, which can be expanded to an astonishing 1M+ tokens using YaRN , optimizing it for deep, repository-scale comprehension. The encoding and decoding steps are also optimized by an upgraded 250k vocabulary, which provides an overall efficiency boost of 10-60% for most global languages.

Performance Evaluation with Other Models

The GPQA (Graduate-level reasoning) benchmark, assessed in the context of the primary language results, is one of the most important measures of the model's ability to reason at a high cognitive level. The performance of Qwen3.5-397B-A17B on this benchmark was remarkable; with a score of 88.4, it significantly exceeded that of Claude 4.5 Opus (87.0) and is highly competitive as compared to other leading models such as Gemini-3 Pro (91.9) and GPT-5.2 (92.4). The GPQA benchmark is critical in validating the quality of the model's Unified Vision-Language Foundation and validating the success of the early fusion training.

source - https://qwen.ai/blog?id=qwen3.5

Within the Vision Language Evaluation space, the MathVision benchmark tests how well models can perform logical reason through visual means with very complex mathematics that requires multi-step operations and reasoning. Qwen3.5-397B-A17B’s 88.6 score on the benchmark dwarfs the scores of Claude 4.5 Opus (74.3) and Gemini 3 Pro (86.6). As such, the model's spatial intelligence is unmatched. This benchmark demonstrates that the model's ability to create very fine-grained pixel-based relationships used to reason logically across multi-step operations rivals even the best dedicated vision models like Qwen3-VL for performing deep spatial and mathematical processing.

source - https://qwen.ai/blog?id=qwen3.5

In addition to the flagship assessments, further evaluation across a wide variety of benchmarks confirms that this model continues to demonstrate dominance. For example, it displayed an impressive ability to retain general knowledge while completing MMLU-Pro and MMLU-Redux assessments and demonstrated an ability to adhere accurately to commands while completing IFEval and IFBench assessments. Agentic tool usage and independent software engineering were validated rigorously via BFCL-V4 and SWE-bench Verified and thus continue to offer significant competition with proprietary systems. Additionally, ultra-long context processing and complex visual hierarchies were validated at the highest level via outstanding performance in Video-MME (video reasoning) and OmniDocBench (document comprehension). Specialized tests such as MedXpertQA-MM, and tests across 201 languages, further demonstrate robust adaptability within niche medical domain and to widely varying global needs.

How to Access and Use Qwen3.5

Qwen3.5 is highly democratized and open-source software under the Apache 2.0 license, which supports both commercial and research-oriented usage. The official API is safely hosted through Alibaba Cloud Model Studio, which is fully compatible with the conventional OpenAI and Anthropic API formats. For users interested in self-hosting the model, weights can be accessed from the Hugging Face repository . It supports frameworks such as vLLM, SGLang, llama.cpp, and MLX. Moreover, developers using large codebases are advised to refer to the official GitHub repository, which contains the Qwen Code CLI open-source terminal agent.

Limitations 

Although Qwen3.5 has made enormous strides, it comes with some operational limitations. The deployment of static YaRN is based on a fixed scaling factor, which may have the unintended consequence of slowing down performance on shorter texts. There is also a minute performance deficit relative to the latest proprietary solutions for managing complex software engineering projects of enormous scale.

Future Work

Future enhancements will continue to work towards developing better user experiences across environments, particularly for navigation by robotic systems, improving autonomous self-improvement through machine learning using environmental feedback loops, and expanding agent-based tasks in the area of cyber-security within the digital world.

Conclusion

If you are an organization creating future systems (hardware clusters, robotic logistics, securing networks or maintaining huge software repositories), the bottom line is going to be not only how rapidly the various models can operate, but also how able those models will be to develop a sustainable means for thinking at the scale required.

Sources:
Blog: https://qwen.ai/blog?id=qwen3.5
GitHub Repo: https://github.com/QwenLM/Qwen3.5
Hugging Face:  https://huggingface.co/Qwen/Qwen3.5-397B-A17B

Disclaimer - This article is intended purely for informational purposes. It is not sponsored or endorsed by any company or organization, nor does it serve as an advertisement or promotion for any product or service. All information presented is based on publicly available resources and is subject to change. Readers are encouraged to conduct their own research and due diligence.