Cyberax AI Playbook
cyberax.com
← All models
Model · Mistral AI · Code

Codestral

Mistral's code-specialised model. Trained on 80+ programming languages; the open-weights default for self-hosted code completion and refactoring.

Modality
Code
License
Mistral Non-Production License (Open weights)
Parameter size
22B
Context window
32,768 tokens
Released
May 29, 2024
Last verified
May 10, 2026
Runs locally
Yes

Strengths

  • Coverage across 80+ programming languages
  • Fill-in-the-middle support — works in IDE completion flows
  • Runs on a single 48GB GPU

Weaknesses

  • Non-commercial weights license — paid tier needed for production
  • Smaller than DeepSeek Coder V2

Try it

WhereTypeNotes
Mistral La Plateforme hosted-api API key required
Hugging Face weights Non-production license
Ollama local ollama run codestral

Used in solutions

Comparison

Cursor vs Copilot vs Claude Code for coding assistance

Official sources

Change log

  • — Initial entry.
  • — Marked deprecated after verification that Codestral 25.08 is the current Codestral release.

Related models

Devstral 2

Mistral AI · Code

Mistral's frontier code-agents model for software engineering tasks. It is designed for tool-heavy coding workflows across whole repositories and multi-file edits.

Open weights Runs locally 256k ctx
Try on: Mistral La Plateforme · Mistral model card
Released Dec 2025 · Verified 2026-05-10

Codestral 25.08

Mistral AI · Code

Mistral's current code-completion model, released at the end of July 2025. It is tuned for low-latency fill-in-the-middle and high-frequency code generation tasks.

Mistral Premier Proprietary 128k ctx
Try on: Mistral La Plateforme
Released Jul 2025 · Verified 2026-05-10

Qwen2.5-Coder 32B

Alibaba · Code

Alibaba's flagship open code model. 32B parameters and Apache 2.0 — the strongest open coding model that fits on a single workstation GPU.

Apache 2.0 Open Runs locally 131k ctx
Try on: Hugging Face · Ollama · OpenRouter
Released Nov 2024 · Verified 2026-05-10

DeepSeek-Coder V2

DeepSeek · Code

DeepSeek's open code model. MoE architecture with strong coverage across 338 programming languages; the open-weights coder of choice for high-end self-hosting.

DeepSeek License Open weights Runs locally 128k ctx
Try on: Hugging Face · DeepSeek Platform · Ollama
Released Jun 2024 · Verified 2026-05-10

StarCoder2 15B

BigCode · Code

BigCode's collaborative code model. 15B parameters trained on 600+ programming languages; strong fit for IDE completion and self-hosted code search.

BigCode OpenRAIL-M Open weights Runs locally 16k ctx
Try on: Hugging Face · Ollama
Released Feb 2024 · Verified 2026-05-10