AMA

Scenario guide

Best AI models for Local / Self-Hosted Coding

A coding assistant you run on your own hardware — no API bills, full privacy, works offline. Restricted to open-weight, self-hostable models and weighted toward practical coding and agentic ability, with a slice for the self-correction that smaller local models lean on inside a good harness. There is no cost axis because inference runs on hardware you already own.

Rankings use the same scenario weights and cost blending as the interactive leaderboard on AI Model Analyzer. Data is min-max normalised per benchmark; missing scores are skipped without penalty.

  1. 1
    DeepSeek R1
    DeepSeek
    Score 86.2Q 86.2In $0.55/M
  2. 2
    DeepSeek V3 (Thinking)
    DeepSeek
    Score 77.9Q 77.9In $0.27/M
  3. 3
    Kimi K2
    Moonshot (Kimi)
    Score 75.0Q 75.0In $0.60/M
  4. 4
    Qwen3 235B
    Alibaba (Qwen)
    Score 73.1Q 73.1In $0.20/M
  5. 5
    Qwen3 235B (Thinking)
    Alibaba (Qwen)
    Score 63.1Q 63.1In $0.20/M
  6. 6
    DeepSeek V3
    DeepSeek
    Score 61.0Q 61.0In $0.27/M
  7. 7
    GLM-4.7
    Zhipu AI (GLM)
    Score 60.4Q 60.4In $0.50/M
  8. 8
    GLM-4.6
    Zhipu AI (GLM)
    Score 60.2Q 60.2In $0.50/M
  9. 9
    Llama 3.3 70B Instruct
    Meta
    Score 58.3Q 58.3In $0.88/M
  10. 10
    Qwen2.5 72B Instruct
    Alibaba (Qwen)
    Score 50.7Q 50.7In $0.90/M
  11. 11
    Llama 3.1 70B Instruct
    Meta
    Score 46.6Q 46.6In $0.88/M
  12. 12
    Llama 4 Scout
    Meta
    Score 24.9Q 24.9In $0.18/M
  13. 13
    Llama 3.1 405B Instruct
    Meta
    Score 20.8Q 20.8In $3.50/M
  14. 14
    Mistral Large 2
    Mistral
    Score 20.3Q 20.3In $2.00/M
  15. 15
    Llama 4 Maverick
    Meta
    Score 19.2Q 19.2In $0.27/M
Open interactive leaderboard Build custom weightsHome