MistralAIProprietary

Mistral Medium 3.5

Name: Mistral Medium 3.5
Author: MistralAI

Compare this model

A high-performance model developed by Mistral AI. It excels in balancing efficiency and performance.

Share:X Hatena LINE

Parameters

Undisclosed

Context Window

License

Proprietary

Release Date

2026-05-01

Japanese Language Capability

✅High-Quality JP

Multilingual model with strong Japanese language processing capabilities.

API Pricing

API pricing for this model is not yet available

Strengths

Weaknesses

Use Cases

Deep Analysis

SWE-Bench Verified

77.6%

Top open-weight model, trails Claude Sonnet 4.6 (79.6%)

τ³-Telecom

91.4%

Best-in-class agentic tool use

Parameters

128B (dense)

All parameters active per token

Context Window

256K tokens

Larger than Claude Sonnet 4.6 (200K)

Input Price

$1.50/1M

Cheaper than GPT-4o ($2.50)

Open Weights

Yes (Modified MIT)

Revenue cap: $20M/month for commercial use

Strengths

・Single unified model replaces three specialized predecessors (instruction, reasoning, coding).
・State-of-the-art open-weight coding agent performance (77.6% SWE-Bench).
・Configurable reasoning effort per request (fast reply vs. deep thinking).

Weaknesses

・High API output pricing ($7.50/1M) relative to some open-weight competitors.
・Self-hosting requires significant hardware (minimum 4x H100 80GB GPUs).
・Incomplete official benchmark reporting (missing MMLU, GPQA, etc.).

Competitor Comparison

Model	Arena	SWE	GPQA	Price
Claude Sonnet 4.6	N/A	79.6%	N/A	$3.00/$15.00
DeepSeek V4 Pro	N/A	80.6%	N/A	$1.74/$3.48
Qwen 3.6 27B	N/A	72.4%	N/A	$0.20/$0.60

Overview

Mistral Medium 3.5 is Mistral AI's April 2026 flagship model, representing a strategic consolidation of their previous specialized models (Magistral for reasoning, Devstral 2 for coding, and Medium 3.1 for instruction-following) into a single, dense 128B parameter architecture. This 'merged model' approach aims to simplify deployment and eliminate the need for model routing in complex agentic workflows. It features a massive 256K token context window, configurable reasoning effort, and a custom-trained vision encoder for handling variable image sizes. Positioned as the premier open-weight coding agent model, Medium 3.5 achieves 77.6% on SWE-Bench Verified, placing it in elite company just behind closed-source leaders like Claude Sonnet 4.6. Its standout agentic capability is demonstrated by a 91.4% score on the τ³-Telecom benchmark. The release is tightly integrated with Mistral's product ecosystem, powering the new asynchronous cloud coding agents in Vibe and Le Chat's Work Mode. While its API pricing is competitive for its capability tier, the Modified MIT license and substantial self-hosting requirements create a specific value proposition for teams needing frontier performance with data sovereignty or high-volume agentic use cases.

Benchmarks & Performance

Detailed Comparison

### vs. Claude Sonnet 4.6 - **Performance**: Claude Sonnet 4.6 leads slightly on SWE-Bench (79.6% vs 77.6%). Both are capable agentic and reasoning models. - **Pricing**: Medium 3.5 is **50% cheaper on input tokens** ($1.50 vs $3.00) and 50% cheaper on output tokens ($7.50 vs $15.00). - **Key Difference**: **Open Weights vs. Closed**. Medium 3.5 can be self-hosted, fine-tuned, and used on-premises for data sovereignty. Claude is API-only. - **Context Window**: Medium 3.5 offers a larger 256K window vs. Sonnet's 200K input / 64K output. - **Best For**: Choose Medium 3.5 for cost efficiency, self-hosting needs, or the Vibe agent workflow. Choose Sonnet for absolute peak coding benchmark performance or pure API simplicity. ### vs. DeepSeek V4 Pro - **Performance**: DeepSeek V4 Pro achieves a higher SWE-Bench Verified score (80.6% vs 77.6%). - **Pricing**: DeepSeek is significantly cheaper, especially on output ($3.48/M vs $7.50/M). - **Key Difference**: **Open Weights (Mistral) vs. Proprietary API (DeepSeek)**. Medium 3.5 allows full self-hosting and customization. - **Best For**: Choose DeepSeek for lowest-cost API access to frontier coding performance. Choose Medium 3.5 for self-hosted deployment, open-weight flexibility, and integrated Vibe agent capabilities. ### vs. Qwen 3.6 27B (MoE) - **Performance**: Medium 3.5 outperforms on SWE-Bench (77.6% vs 72.4%) and is a much larger, more capable model overall. - **Pricing & Efficiency**: Qwen 3.6 is radically cheaper (Apache 2.0, ~$0.20/$0.60 per million tokens) and runs on consumer GPUs. - **Key Difference**: **Unified 128B Dense vs. Efficient 27B MoE**. Medium 3.5 is for high-capability, long-horizon agentic workloads. Qwen 3.6 is for cost-sensitive, general-purpose coding tasks with lower infrastructure requirements. - **Best For**: Choose Qwen 3.6 for budget-conscious development or personal use. Choose Medium 3.5 for enterprise agentic workflows, maximum coding capability, and the ability to scale reasoning per task.

Community Feedback

Reception among developers and researchers is positive but measured, with clear distinctions drawn based on use case: - **Coding Agent Enthusiasts**: Highly positive about the 77.6% SWE-Bench score and the practical 'merged model' concept. The Vibe remote agents (async PR generation) are highlighted as a genuine workflow innovation. ([Awesome Agents Review](https://awesomeagents.ai/reviews/review-mistral-medium-3-5/)) - **Enterprise/Architect View**: Appreciate the consolidation from three models to one, simplifying deployment. The open-weight nature is critical for on-premises and data-sensitive use cases, despite the Modified MIT license's revenue cap. - **HackerNews & Developer Forums**: Sentiment is more critical. Common themes include: 1) Benchmark cherry-picking (lack of GPQA/MMLU data), 2) The 'efficiency paradox' of a dense 128B model vs. more efficient MoE alternatives, 3) Sticker shock at the 3.75x price increase over its predecessor, Mistral Medium 3.1, and 4) Concerns about the practical speed of self-hosted quantized versions on non-data-center hardware. ([ChatForest Community Section](https://chatforest.com/reviews/mistral-medium-3-5-dense-128b-agentic-llm-review/)) - **Adoption Pattern**: Early adoption is strongest among: a) Teams already in the Mistral ecosystem using Le Chat or Vibe, b) Organizations requiring self-hosted frontier models for coding/agentic tasks, and c) Developers valuing the unified endpoint for mixed instruction-reasoning-coding agents.

Use Cases

1. **Agentic Coding with Vibe Remote Agents**: The primary use case. Deploy teams of asynchronous coding agents that can work on module refactors, test generation, and bug fixes in parallel, offloading work to the cloud and submitting GitHub PRs for review. Choose this over Claude or DeepSeek when you need the integrated Vibe/Le Chat workflow and the ability to 'teleport' sessions from CLI to cloud. 2. **Enterprise AI with Data Sovereignty**: For regulated industries (finance, healthcare, defense) or companies with strict data governance, the model can be self-hosted on-premises or in a private cloud (requires 4+ H100 GPUs). Choose this over all proprietary models (Claude, GPT) when data cannot leave your perimeter. 3. **Unified Multi-Task API Endpoint**: Replace a suite of specialized models with a single API call. Use `reasoning_effort="none"` for fast chat and simple tool calls, and `reasoning_effort="high"` for complex code generation, mathematical reasoning, or multi-step planning. This simplifies architecture for applications that dynamically mix task types. 4. **Multimodal Document & Code Analysis**: Leverage the custom vision encoder to process varied visual inputs—architectural diagrams, UI screenshots, mobile app interfaces, or documents with complex layouts—without forced resizing. Ideal for visual debugging, documentation generation, or extracting data from mixed-media sources.

Latest News

**Release Date**: April 29, 2026, with simultaneous product launches: - **Vibe Remote Agents**: Asynchronous, cloud-based coding agents that can open GitHub PRs autonomously. Session 'teleportation' from local CLI to cloud is a key feature. - **Le Chat Work Mode**: A new agentic mode for complex, multi-step tasks across connected tools (email, calendar, Jira, etc.). - **Pricing & License**: Set at $1.50/$7.50 per million input/output tokens. Released under a **Modified MIT License** with a revenue cap for commercial use (companies with >$20M/month revenue need a separate license). - **Infrastructure**: Day-zero support for inference via **vLLM** and **SGLang**. An **EAGLE** speculative decoding model was released to accelerate local inference. - **Partner Integrations**: Available for prototyping on **NVIDIA build.nvidia.com** and as a production container via **NVIDIA NIM**. **Note on Context**: The model has been in public preview since late April 2026. Mistral has indicated that availability on major cloud providers (AWS Bedrock, Azure AI, Google Vertex) is planned but not yet confirmed as of the latest reports.

Positioned as the premier open-weight coding agent model, Medium 3.5 achieves 77.6% on SWE-Bench Verified, placing it in elite company just behind closed-source leaders like Claude Sonnet 4.6. Its standout agentic capability is demonstrated by a 91.4% score on the τ³-Telecom benchmark. The release is tightly integrated with Mistral's product ecosystem, powering the new asynchronous cloud coding agents in Vibe and Le Chat's Work Mode. While its API pricing is competitive for its capability tier, the Modified MIT license and substantial self-hosting requirements create a specific value proposition for teams needing frontier performance with data sovereignty or high-volume agentic use cases.

Sources

Analysis generated: 2026-07-17