Mistral AI shipped Mistral Medium 3.5 on June 5, 2026. The open-weight model drops as an Apache 2.0 release with 120 billion parameters, a 128,000 token context window, and 140GB of downloadable weights. The announcement appeared on the Mistral AI homepage and on Hugging Face, where the model card lists benchmark results, quantization options, and deployment guidance. Unlike the earlier proprietary Medium API model, this version is meant for self-hosting, fine-tuning, and commercial use without per-token fees. The release lands in a crowded open-source field, including recent launches from Qwen, DeepSeek, Meta, and Cohere. Free AI readers watching paid API pricing changes should see this as a meaningful shift toward capable models that do not require a credit card.

Mistral Medium 3.5 matters because it narrows the gap between open-weight and closed models. The model card reports MMLU at 84.2, HumanEval at 78.9, and GPQA at 48.7. Those scores sit close to several paid flagship APIs from early 2026, but with a license that lets teams run the model on their own hardware. For developers burned by the latest round of API free tier cuts and usage-based billing changes, this release offers a real exit ramp. You can compare it with other open models in our best open-source LLM models 2026 guide.

Mistral AI has been aggressive on open releases, but Medium 3.5 is different from Mistral Small 4 and the older Mixtral family. It is a dense model, not a mixture of experts, which makes it easier to run with standard transformers, vLLM, and llama.cpp. The vendor positions it for coding agents, RAG pipelines, and local data analysis. The release date of June 5, 2026 follows a wave of pricing changes across OpenAI, Anthropic, and Google, all covered in our major AI model tier changes tracker.

Why now? Open-weight models are becoming the default for self-hosted teams that cannot tolerate API rate limits or surprise billing. Mistral Medium 3.5 packs enough reasoning capability to handle long documents and multi-step tool calls without a frontier API. The weights are available through Mistral AI and Hugging Face, and early community benchmarks show strong performance on agentic coding tasks. This guide breaks down the technical specs, license, hardware requirements, and how the model fits into the free AI landscape. For a focused look at the release, see our Mistral Medium 3.5 open-weight guide.

How Do the Top Options Compare?

Model Best For Parameters Context Window License Weight Size
Mistral Medium 3.5 Agentic coding and long documents 120B dense 128k tokens Apache 2.0 140GB fp16
Mistral Small 4 Fast local tasks 44B dense 128k tokens Apache 2.0 52GB fp16
Llama 4 Scout Extreme long context and multilingual edge 109B MoE 17B active 10M tokens Llama 4 Community 220GB fp16
Qwen 3.6 Apache Open-source coding and math benchmarks 30B MoE 3B active 256k tokens Apache 2.0 34GB fp16
DeepSeek V4 Frontier open math and coding 1.6T MoE 32B active 128k tokens MIT 3.2TB fp16

Weight sizes are approximate fp16 values. Quantized 4-bit versions reduce disk and memory use by 60 to 70 percent.

1. Mistral Medium 3.5 , Agentic coding and long documents

Mistral Medium 3.5 is a dense 120 billion parameter model with a 128,000 token context window. The weights weigh 140GB in fp16 and are available on Hugging Face under Apache 2.0. The model card reports MMLU 84.2, HumanEval 78.9, GPQA 48.7, and a 20 percent lower hallucination rate on long-context retrieval tasks than Mistral Small 4. That mix makes it a practical middle ground between small local models and paid frontier APIs. The license is the headline. Apache 2.0 means you can fine-tune the model, deploy it in commercial products, and modify it without opening your own code. That is a direct contrast to some competing open-weight releases that carry restrictive community licenses. For teams already tired of API billing changes, like the Anthropic agent billing split and GitHub Copilot usage-based billing, the ability to self-host removes per-token charges entirely. Running Mistral Medium 3.5 requires real hardware. A single 80GB GPU can serve a 4-bit quantized version, while fp16 needs about 160GB of VRAM across two or more accelerators. CPU inference is possible with llama.cpp, but expect slow token generation. The model supports function calling, JSON mode, and a 128k context window that works well for long documents, agentic coding, and RAG. See our open-source LLM self-host guide for setup options.

Key strengths:

  • ✅ Apache 2.0 license permits commercial use, fine-tuning, and private deployment
  • ✅ 128k token context handles long codebases and document analysis
  • ✅ Dense architecture simplifies serving with vLLM, TensorRT-LLM, and llama.cpp
  • ✅ Benchmarks close the gap with paid models for coding and reasoning
  • ✅ Function calling and JSON mode support agentic workflows
  • ❌ 140GB fp16 weights demand high-end GPUs or multiple accelerators
  • ❌ Not as fast as smaller models like Mistral Small 4 for simple tasks
  • ❌ Fewer community quantizations at launch than older model families

Who it’s for: Developers and small teams that need a self-hosted, commercially safe model for long-context coding and agentic tasks without per-token API fees.

2. Mistral Small 4 , Fast local coding and RAG

Mistral Small 4 is the smaller sibling in Mistral AI’s open-weight lineup. It is a 44 billion parameter model with a 128k token context window and Apache 2.0 license. The 52GB fp16 footprint fits on a single 80GB GPU, and 4-bit quantizations run comfortably on 24GB cards. That makes it the easiest Mistral model to deploy on local hardware for coding assistants and RAG pipelines. Benchmarks are lower than Medium 3.5 but strong for the size. Mistral Small 4 scores 78.1 on MMLU, 71.3 on HumanEval, and handles 60k tokens of retrieval accuracy with minimal degradation. The model is fast, often generating 50 to 70 tokens per second on an A100. For users comparing open alternatives, see our Mistral Small 4 Apache open-source guide and the latest Mistral AI open-source release tracker. Mistral Small 4 makes sense when latency and cost matter more than peak reasoning. It lacks the deep multi-step reasoning of Medium 3.5, but it still supports function calling and fine-tuning. The Apache license means no restrictions for commercial products. If you are watching the shift away from free API tiers, this model is a low-cost way to keep a local fallback for coding and document tasks.

Key strengths:

  • ✅ Small enough to run on a single 24GB GPU with 4-bit quantization
  • ✅ Apache 2.0 license is safe for production and commercial use
  • ✅ Faster inference than Medium 3.5 on the same hardware
  • ✅ Low VRAM requirements make it ideal for local coding tools
  • ❌ Lower benchmark scores than Medium 3.5 on complex reasoning and math
  • ❌ Struggles with very long agentic workflows beyond 128k tokens
  • ❌ Smaller model capacity limits fine-tuning on niche domain data

Who it’s for: Teams that need a fast, cheap, local model for everyday coding and document tasks without high-end GPU hardware.

3. Llama 4 Scout , Extreme long context and multilingual edge

Meta AI’s Llama 4 Scout is a mixture-of-experts model with 109 billion total parameters and 17 billion active parameters. Its headline feature is a 10 million token context window, far beyond the 128k of Mistral Medium 3.5. The weights are available through Meta AI and Hugging Face under the Llama 4 Community License. That license allows research and commercial use but imposes restrictions for very large deployments. For long documents, code repositories, and multi-day agent sessions, Scout is the strongest open-weight option. It scores 82.4 on MMLU and performs well on multi-hop retrieval at 100k and 1M token depths. However, the 220GB fp16 size and MoE routing make it harder to serve than dense models. Our Llama 4 Scout and Maverick open-source guide covers setup and benchmarks. Scout is not a direct replacement for Mistral Medium 3.5 because of licensing and hardware trade-offs. The Community License restricts use by products with more than 700 million monthly active users, and the 10M context demands aggressive KV cache management. But for teams processing huge codebases or long legal documents, the trade-off may be worth it.

Key strengths:

  • ✅ 10 million token context window handles massive codebases and document sets
  • ✅ Active parameter efficiency reduces inference compute compared with dense 109B
  • ✅ Strong multilingual performance across 12 languages
  • ✅ Available through Meta AI and Hugging Face with community support
  • ❌ Llama 4 Community License has deployment restrictions for very large products
  • ❌ 220GB fp16 size plus MoE routing complicates single-GPU serving
  • ❌ 10M context windows require substantial KV cache memory and software tuning

Who it’s for: Teams that need extreme long-context retrieval and multilingual capabilities and can accept Meta’s community license terms.

4. Qwen 3.6 Apache , Open-source coding and math benchmarks

Qwen 3.6 Apache is Alibaba’s open-weight release designed for coding, math, and tool use. It has 30 billion total parameters with 3 billion active parameters, so inference is surprisingly cheap. The model uses a 256k token context window and Apache 2.0 license. It runs on a single 24GB GPU at 4-bit quantization and scores 83.1 on MMLU, 84.5 on HumanEval, and 68.2 on GPQA, often beating larger dense models. The small active parameter count gives Qwen 3.6 Apache strong throughput for agentic coding. It supports function calling, parallel tool calls, and long code generation. For developers who prioritize benchmark performance and permissive licensing, this is one of the best open models in 2026. See our Qwen 3.6 Apache open-source coding guide for full results and deployment examples. Compared with Mistral Medium 3.5, Qwen 3.6 Apache is smaller and easier to run, but it may not hold reasoning depth across very long 128k to 256k documents. The Apache 2.0 license and low hardware requirements make it a favorite for local coding assistants and CI pipelines. It is also a reminder that Alibaba continues to push open models while Google cuts Gemini API prices and paid tiers shift.

Key strengths:

  • ✅ Apache 2.0 license with unrestricted commercial use
  • ✅ 30B total, 3B active parameters runs on a single 24GB GPU
  • ✅ 256k token context window exceeds Mistral Medium 3.5
  • ✅ Top-tier coding and math benchmarks for its size
  • ❌ Smaller capacity may underperform on complex multi-step reasoning at long depths
  • ❌ Less established community support than Mistral and Meta models
  • ❌ Hardware benchmarks vary widely across quantization backends

Who it’s for: Developers who want the best permissive open coding model on modest local hardware and do not need 120B scale reasoning.

5. DeepSeek V4 , Frontier open math and coding

DeepSeek V4 is a massive mixture-of-experts model with 1.6 trillion total parameters and 32 billion active parameters. It uses a 128k token context window and MIT license. The model made waves by matching GPT-5.5 level reasoning on math and code benchmarks while allowing free commercial use. You can find it on DeepSeek and Hugging Face. DeepSeek V4 scores 91.2 on MATH, 88.7 on HumanEval, and 69.4 on GPQA. The active parameter count keeps inference cost lower than dense models of similar total size, but serving 1.6T parameters still requires serious cluster hardware. Our DeepSeek V4 open-source guide covers deployment, quantization, and benchmarks in detail. For teams that need best-in-class open reasoning and do not mind the infrastructure burden, DeepSeek V4 is the strongest option in 2026. It trades the moderate hardware requirements of Mistral Medium 3.5 for higher peak math and code performance. The MIT license is even more permissive than Apache 2.0, removing most attribution requirements.

Key strengths:

  • ✅ MIT license is extremely permissive for commercial and research use
  • ✅ Top open benchmarks for math and coding with 91.2 MATH and 88.7 HumanEval
  • ✅ MoE active parameters keep inference cheaper than dense 1.6T models
  • ✅ Strong agentic tool calling and long reasoning traces
  • ❌ 1.6T total parameters requires multi-node GPU clusters for full precision
  • ❌ High operational complexity compared with 120B dense models
  • ❌ Context window capped at 128k, shorter than Qwen and Llama Scout

Who it’s for: Research labs and enterprises that need frontier-level open reasoning and can manage large-scale GPU infrastructure.

Frequently Asked Questions

What is Mistral Medium 3.5?

Mistral Medium 3.5 is a 120 billion parameter open-weight language model from Mistral AI. It has a 128,000 token context window, Apache 2.0 license, and ships with 140GB of fp16 weights. It targets coding, agentic workflows, and long document reasoning.

Is Mistral Medium 3.5 free to use commercially?

Yes. The Apache 2.0 license allows commercial use, modification, and private deployment. You do not owe Mistral AI per-token fees or royalties when you self-host or fine-tune the model. Standard liability and trademark conditions still apply.

How much GPU memory do I need to run Mistral Medium 3.5?

A 4-bit quantized version runs on a single 80GB GPU. FP16 inference needs about 160GB of VRAM, often split across two A100 or H100 cards. CPU inference works but is very slow for interactive use.

How does Mistral Medium 3.5 compare to paid models like GPT-5.5?

Early benchmarks show Mistral Medium 3.5 scoring 84.2 on MMLU and 78.9 on HumanEval, which places it near several paid flagship models from early 2026. The difference is that self-hosting removes per-token API costs and usage caps. Latency and serving quality depend on your hardware.

When did Mistral Medium 3.5 release?

Mistral AI released Mistral Medium 3.5 on June 5, 2026. The company announced the release on its homepage and on Hugging Face.

Can I fine-tune Mistral Medium 3.5?

Yes. The Apache 2.0 license permits full fine-tuning, LoRA, and distillation. You can adapt the model on your own data without opening your source code. Fine-tuning large 120B models still requires substantial GPU memory.

What Should You Remember?

  • Mistral Medium 3.5: dense 120B parameter model with 128k token context and 140GB fp16 weights.
  • License: Apache 2.0, so commercial use, fine-tuning, and private deployment are allowed.
  • Benchmarks: MMLU 84.2, HumanEval 78.9, GPQA 48.7, close to paid frontier models.
  • Hardware: 80GB GPU for 4-bit, about 160GB VRAM for fp16 across multiple accelerators.
  • Release date: June 5, 2026, announced on Mistral AI and Hugging Face.
  • Why it matters: self-hosted model avoids per-token API pricing and usage limits.
  • Alternatives: Mistral Small 4, Llama 4 Scout, Qwen 3.6 Apache, and DeepSeek V4 each trade speed, context, or license.

Free AI News is an independent editorial publication. Information about AI pricing, free-tier limits, and features changes frequently and may become outdated. Always verify current details through the vendor’s official pages. Affiliate links may earn a commission at no cost to you, and never affect our reporting.