Agents Directory
SkillsRankingsAgents
CategoriesModelsBenchmarksCompareAgent LeaderboardSkillsRankingsAgentsAbout
/Rankings
/Best open-source AI models

Best open-source AI models

Open source

Open-weight models you can download, inspect, and self-host now sit within reach of the closed frontier. These are the best open-source AI models in 2026, ranked on real capability, from full-size flagships to models you can run on a workstation.

Which model should you run?

As of June 2026, the right open-source model depends on where you run it:

  • Hosted via API: MoonshotAIKimi K2.6 or DeepSeekDeepSeek V4, the open frontier at a fraction of closed prices
  • Coding agents: Qwen3 Coder 480B for big repositories, Devstral 2 if you want something self-hostable on smaller hardware
  • Your own servers: OpenAIgpt-oss-120b and NvidiaNemotron 3 Ultra are the practical self-host flagships; Llama 4.1 405B if you have the GPUs
  • A workstation or home lab: see our best models for Ollama ranking for what actually fits in consumer hardware

The best open-source AI model right now is MoonshotAIKimi K2.6 from Moonshot AI: the strongest open-weight agentic coder we track, at 47.6% on CursorCursorBench 3.1, with a 262K context window and a free hosted tier on OpenRouter.

DeepSeekDeepSeek V4 is the close second, pairing near-frontier reasoning with a 1M token context window at $0.435 per million input tokens hosted, and GLM 5.1 logoGLM 5.1 rounds out the open frontier. For coding specifically, Qwen3 Coder 480B is the specialist with a 1M context window.

The headline of 2026: open-weight models are no longer a compromise. The top picks here compete directly with closed flagships on agent work while staying inspectable, hostable, and far cheaper.

We also have more in-depth rankings:Best models to run with Ollama
The ranking
Updated June 2026
Best open-source models overall

The open frontier: flagship-class capability with weights you can download.

#ModelContextInput
  • 1
    1MoonshotAI
    Kimi K2.6Moonshot AIThe strongest open-weight agentic coder on our board (47.6% on CursorBench).
    Context262K
    Input$0.66/M
  • 2
    2DeepSeek
    DeepSeek V4DeepSeekOpen weights, near-frontier reasoning, and a 1M token context window.
    Context1.049M
    Input$0.435/M
  • 3
    3GLM 5.1 logo
    GLM 5.1Z.AIZ.AI's flagship GLM: a strong open-weight all-rounder for agent work.
    Context203K
    Input$0.98/M
Best open-source for coding

Specialists tuned for repositories, agentic edits, and long tool chains.

#ModelContextInput
  • 1
    1Qwen
    Qwen3 Coder 480B A35BQwenThe open coding specialist: 1M context for large repositories, free tier included.
    Context1.049M
    Input$0.22/M
  • 2
    2Mistral
    Devstral 2 2512Mistral AIMistral's open coding model, sized to self-host sensibly.
    Context262K
    Input$0.4/M
Best from US labs

The open-weight releases from Nvidia, OpenAI, and Meta.

#ModelContextInput
  • 1
    1Nvidia
    Nemotron 3 UltraNvidiaNvidia's open-weight flagship with a 1M context window, free on OpenRouter.
    Context1M
    Input$0.5/M
  • 2
    2OpenAI
    gpt-oss-120bOpenAIOpenAI's open-weight release: a dependable generalist you can self-host.
    Context131K
    Input$0.039/M
Community and value picks

Lightly aligned and budget-friendly open models with loyal followings.

#ModelContextInput
  • 1
    1Minimax
    MiniMax M2.5MiniMaxOpen weights and solid quality at $0.15 per million when hosted.
    Context205K
    Input$0.15/M
  • 2
    2NousResearch
    Hermes 3 405B InstructNous ResearchNous Research's lightly aligned large model, a community favorite.
    Context131K
    Input$1/M
Frequently asked questions
What is the best open-source AI model in 2026?

MoonshotAIKimi K2.6 from Moonshot AI is the best open-source model right now: the strongest open-weight agentic coder we track at 47.6% on CursorCursorBench 3.1. DeepSeekDeepSeek V4 is the close second with near-frontier reasoning and a 1M token context window, and GLM 5.1 logoGLM 5.1 from Z.AI rounds out the open frontier.

Are open-source models as good as GPT-5.5 or Claude?

On agentic coding benchmarks the gap has narrowed dramatically: MoonshotAIKimi K2.6 scores 47.6% on CursorCursorBench versus the 60s for closed flagships like OpenAIGPT-5.5 and ClaudeClaude Opus 4.8. For most everyday work the difference is hard to notice, and the open models cost 5 to 20 times less. The closed frontier still wins on the hardest multi-step reasoning.

What is the best open-source model for coding?

MoonshotAIKimi K2.6 for agentic coding overall, and Qwen3 Coder 480B as the dedicated specialist with a 1M token context window for large repositories. Devstral 2 from Mistral is the best pick if you want an open coding model you can realistically self-host on modest hardware.

Can I use open-source models commercially?

Usually yes, but check each license. Many open-weight models ship under permissive licenses, while some (like certain Llama releases) add usage conditions. The weights being downloadable does not automatically mean unrestricted commercial use, so read the license on the model's repository before shipping.

Share:
Details:

  • Models

    10

  • Filter

    Open source

  • Updated

    June 2026
Browse:SkillsRankingsModelsBenchmarksProvidersAgentsAgent LeaderboardCompareCategories
Quick Links:AboutBlog

© 2026 Agents Directory