Mistral Medium 3.5
Mistral AIMULTIMODAL Open weight
- Provider
- Mistral AI
- Modality
- MULTIMODAL
- Parameters
- 128B
- Context window
- 256,000 tokens
- Weights
- Open
- Released
- 29 Apr 2026
Mistral Medium 3.5 is a 128B dense all-rounder that fuses instruct, reasoning and coding into one self-hostable model aimed at agentic workflows.
Best for
- Unified weights for instruct + reasoning + coding with tunable effort
- Tuned for agentic and coding workflows (ships a PR-opening coding agent)
- Cheaper to self-host than Large 3 (128B dense vs 675B MoE)
- Strong SWE-bench coding for its size
How it compares — and the India angle
- 128B dense self-hosts on a multi-GPU server — cheaper infra than the 675B Large 3.
- Cost-sensitive Indian users get better value from Small 4 or self-hosted Magistral.
- Strong agentic/SWE coding makes it a managed-API option for Indian dev teams without 675B infra.
Benchmarks
Representative public scores (approximate, higher is better) for relative comparison. Check the provider for the latest official results.
SWE-bench Verified78
How to access
API ~$1.50 / 1M input, ~$7.50 / 1M output. Open-weight (Modified MIT) — free to self-host.
Access Mistral Medium 3.5Frequently asked questions
What is Mistral Medium 3.5?
Mistral Medium 3.5 is a multimodal AI model developed by Mistral AI with 128 billion parameters.
Who made Mistral Medium 3.5?
Mistral Medium 3.5 was created by Mistral AI.
Is Mistral Medium 3.5 open source?
Yes, Mistral Medium 3.5 has open weights and can be downloaded and self-hosted.
What is Mistral Medium 3.5's context window?
Mistral Medium 3.5 supports a context window of 256,000 tokens.