LANGUAGE MODEL Mistral AI Last updated:

Mistral Medium 3.5 NEW

Unified Open-Weights Coding Agent

Mistral's flagship dense (non-MoE) 128B model, released under Apache 2.0 with a 256K context window. It folds three previously-separate Mistral specialists — Medium 3.1 (instruction-following), Magistral (reasoning), and Devstral 2 (coding) — into one weight set with a mode toggle. Ships with Vibe, Mistral's remote-agent CLI for autonomous coding tasks.

Try Le Chat API Docs ↗
Speed
Top 10%
154 tok/s output
Cost
High
$1.50 in / $7.50 out
Context
262K
Up to 262,144 tokens
How are Intelligence, Speed & Cost bucketed?
Intelligence and Speed buckets are percentile ranks on Artificial Analysis. Cost buckets are fixed dollar thresholds keyed off output-token price ($/M out).
Intelligence
  • Top 1%≤ 1%
  • Top 5%≤ 5%
  • Top 10%≤ 10%
  • Good≤ 25%
  • Medium≤ 50%
  • Below avg> 50%
Speed
  • Top 1%≥ 345 tok/s
  • Top 5%≥ 237 tok/s
  • Top 10%≥ 196 tok/s
  • Good≥ 146 tok/s
  • Medium≥ 90 tok/s
  • Slow< 90 tok/s
Cost
  • Freeopen weights · self-host
  • Low< $1 / M out
  • Moderate$1–5 / M out
  • High≥ $5 / M out
Official ↗ GitHub ↗ Hugging Face ↗

Why it matters

Closes the gap to Claude Sonnet 4.6 on SWE-Bench Verified (77.6% vs ~80%) with weights you can download. The first time a permissively-licensed model has been a credible drop-in for closed coding agents in regulated deployments. Also notable for being dense at 128B at a moment when almost every other frontier release is MoE — a deliberate bet that dense models are easier to fine-tune and serve in mixed workloads.

Core Capabilities

Long Documents
Handles entire codebases, books, and multi-doc RAG.
Coding
Strong real-world software engineering.
Agent Workflows
Built for tool use and autonomous tasks.

Context Window

262k tokens
≈ 202+ pages
4k Chat 聊天
32k Long docs 长文档
128k Books 整本书
400k Multi-doc 多文档
1M Codebase 整个代码库
10M
262k

Availability

API
Available
Product / App
Not available
Open Source
Released
Enterprise
Contact sales

Pricing Model

Free / self-host
Open weights — pay only for compute.
Self-host

Capability / Performance

Where this model sits relative to the middle 60% of models in the tree. All scores are 0–10 (higher is better).

Lower 20% Upper 80% This model
Reasoning
AA Intelligence Index · scaled to 10
1.7
5.6
5.6
Coding
SciCode · scaled to 10
1.8
4.3
4.0
Agentic tasks
Terminal-Bench Hard · scaled to 10
0.2
3.6
3.3
Context / memory
Context window size · log-scaled
6.0
9.0
7.1
Cost efficiency
Input price ($/M tokens) · cheaper scores higher
6.2
10.0
7.0
Lower 20% 20th percentile — 20% of models score below this This model Where the current model lands Upper 80% 80th percentile — only 20% of models score above this Percentile boundaries are computed across every model in the tree that reports the underlying benchmark for each capability.

What it feels like

  • Reasoning is mixed; expect retries on hard tasks.
  • Coding help is basic; prefer a code-specialized model.
  • Not ideal for autonomous tool-use; keep it tightly scoped.
  • Comfortable with long context — large docs and codebases fit.
  • Good cost/performance; efficient for high-volume workloads.

Best use cases

  • Mistral API (Vibe coding agent) (Mistral AI)
  • Self-hosted enterprise deployments (Mistral AI)

Tools to try

Le Chat La Plateforme Continue.dev Aider

Not ideal for

  • Turnkey hosted reliability (you’ll need deployment/ops).
  • High-fidelity media generation (image/video/audio-first models).