Publication date: May 1, 2026
Mistral AI has released Mistral Medium 3.5, a 128-billion parameter dense language model with a 256,000-token context window and open weights under a modified MIT license. The model handles instruction-following, reasoning, coding, and vision tasks in a single set of weights, scoring 77.6% on SWE-Bench Verified. Mistral Medium 3.5 is available through the Mistral API at $1.50 per million input tokens and $7.50 per million output tokens, as well as on Hugging Face for self-hosting.
What is Mistral Medium 3.5?
Mistral Medium 3.5 is a dense transformer model with 128 billion parameters. Unlike mixture-of-experts architectures used in some competing models, Medium 3.5 uses a single dense set of weights that handles instruction-following, reasoning, coding, and multimodal vision tasks. The model includes a custom vision encoder trained from scratch to process variable image sizes and aspect ratios.
The context window extends to 256,000 tokens, covering both input and output. Reasoning effort is configurable per request, allowing the same model to handle quick conversational replies or work through complex multi-step agentic tasks. Mistral describes it as their “first flagship merged model,” consolidating capabilities that previously required separate specialized models.
Mistral Medium 3.5 benchmarks and performance
On SWE-Bench Verified, a benchmark that tests whether a model can resolve real GitHub issues by generating correct patches, Mistral Medium 3.5 scores 77.6%. This places it ahead of Devstral 2 and Qwen 3.5 397B A17B, and close to Gemini 3.1 Pro Preview, which leads at 78.8%. On the Tau-3 Telecom benchmark for agentic tool-use tasks, the model scores 91.4%, demonstrating strong performance in domain-specific scenarios that require multi-step tool calling.
The model also powers Mistral’s Vibe remote coding agents, replacing Devstral 2 as the default model in the Vibe CLI. In Mistral’s “Work mode” within Le Chat, Medium 3.5 handles async cloud-based coding tasks with structured output for downstream pipelines.
Mistral Medium 3.5 pricing and availability
Mistral Medium 3.5 is priced at $1.50 per million input tokens and $7.50 per million output tokens through the Mistral API. This positions it in a middle tier: more expensive than the budget-oriented Mistral Medium 3 ($0.40/$2.00 per million tokens) but less costly than GPT-4o and Claude Sonnet 4 at comparable parameter counts.
The model is accessible through several channels. It serves as the default model in Le Chat at chat.mistral.ai, with advanced features available on Pro, Team, and Enterprise plans. The API endpoint uses the model identifier “mistral-medium-3.5” with “mistral-medium-3” as an alias. Open weights are published on Hugging Face at mistralai/Mistral-Medium-3.5-128B under a modified MIT license, and the model is also available through NVIDIA’s build.nvidia.com platform and as an NVIDIA NIM containerized inference microservice.
How does Mistral Medium 3.5 compare to GPT-5.5 and Claude Opus 4.7?
Mistral Medium 3.5 enters a competitive landscape alongside OpenAI’s GPT-5.5, released April 23, and Anthropic’s Claude Opus 4.7, released April 16. While GPT-5.5 and Claude Opus 4.7 are closed-source models with higher price points, Medium 3.5 differentiates itself through open weights and self-hosting capability on as few as four GPUs.
On SWE-Bench Verified, Medium 3.5’s 77.6% score is competitive but sits below the top frontier models. Its primary appeal lies in the combination of strong coding and agentic performance with open-weight availability, a 256K context window, and configurable reasoning effort at a moderate price point.
Mistral Medium 3.5 is available now through the Mistral API, Le Chat, and Hugging Face. Full technical documentation is available at docs.mistral.ai/models/model-cards/mistral-medium-3-5-26-04, and the official announcement can be found at mistral.ai/news/vibe-remote-agents-mistral-medium-3-5.
