Mistral Medium 3.5
mistral.ai
April 29, 2026
6 min read
🔥🔥🔥🔥🔥
62/100
Summary
Mistral Medium 3.5 introduces remote coding agents in Vibe, allowing users to run coding tasks in the cloud independently and in parallel. Users can initiate these agents from the Mistral Vibe CLI or within Le Chat, enabling task offloading without leaving the conversation.
Key Takeaways
- Mistral Medium 3.5 is a new 128B dense model that integrates instruction-following, reasoning, and coding capabilities, now available in public preview.
- Remote coding agents in Mistral Vibe can run asynchronously in the cloud, allowing users to offload tasks and continue working without being the bottleneck.
- The new Work mode in Le Chat utilizes Mistral Medium 3.5 to manage complex multi-step tasks, enabling parallel tool calls until completion.
- Mistral Medium 3.5 scores 77.6% on SWE-Bench Verified and 91.4 on τ³-Telecom, demonstrating strong real-world performance and agentic capabilities.
Community Sentiment
Positives
- Mistral Medium 3.5 can run at Q4 with only 70GB of VRAM, making it accessible for consumer-level hardware, which could democratize AI usage.
- The model competes well despite its size, indicating that smaller models can still hold their ground against larger counterparts in certain tasks.
- This model's design allows for multiple instances to run on a single H100 cluster, which is advantageous for enterprises facing VRAM constraints.
Concerns
- Reports indicate that the model's performance is limited, with only about 3 tokens per second on high-end hardware, raising concerns about its practical usability.
- There is a noticeable gap in capability between frontier models and others, suggesting that Mistral may not meet the productivity needs of users accustomed to top-tier models.
- The model struggles with certain tasks, such as drawing SVGs, which highlights its limitations in specific applications compared to other LLMs.