Best open-source / self-hostable LLM for coding (2026)
Open-weight models you can run on your own hardware — your code never leaves your infrastructure. Ranked by SWE-bench among models with open weights or self-host options.
🏆 Top pick: MiMo V2.5 Pro
MiMo V2.5 Pro is the strongest open-weight model we track for coding (79% SWE-bench) — run it locally for full privacy.
The ranked list
| # | Model | SWE-bench | Context window | Hosting | Input price |
|---|---|---|---|---|---|
| 1 | MiMo V2.5 Pro | 79% | 1M+ | Open-weights | Free (self-hosted) |
| 2 | Hy3 Preview | 74% | 256K | Open-weights | $0.30 |
| 3 | Llama 4 Behemoth | 74% | 1M+ | Open/Self-host | $3 |
| 4 | Ling 2.6 (1T) | 72% | 256K | Open-weights | Free (self-hosted) |
| 5 | Qwen 3.5 397B-A17B | 71% | 256K | Open-weights | Free (self-hosted) |
| 6 | Kimi K3 | 70% | 2M | Closed/API + Self-host | $0.60 |
| 7 | Laguna XS.2 | 68% | 131K | Open-weights | Free (self-hosted) |
| 8 | Qwen 3.6 27B | 68% | 256K | Open-weights | Free (self-hosted) |
Why each made the list
1 MiMo V2.5 Pro
Highest open-weight coding performance, 1M context agentic tasks, complex multi-step engineering, long-context reasoning
2 Hy3 Preview
Agent-led tasks in development environments, CodeBuddy/WorkBuddy workflows, low-latency production coding (54% TTFT reduction), open-weight frontier alternative
3 Llama 4 Behemoth
Self-hosted frontier reasoning, complex agentic coding, multimodal analysis
4 Ling 2.6 (1T)
Open-source SOTA execution-heavy tasks, enterprise agent workflows, production coding with optimized token efficiency, AIME-level reasoning
5 Qwen 3.5 397B-A17B
Frontier open-weight reasoning, native multimodal tasks (text/image/video co-trained), commercial use under Apache 2.0, Alibaba Cloud Model Studio (1M ctx hosted)
6 Kimi K3
Agentic coding at low cost, ultra-long context, China-region deployments
7 Laguna XS.2
Local agentic coding on Mac/laptop (runs on 36GB), SWE-bench tasks, long-horizon autonomous coding, Zed/JetBrains integration via ACP
8 Qwen 3.6 27B
Single-GPU agentic coding (fits on 1x H100), workstation deployment, beats much larger MoE models on agentic tasks, Apache 2.0 commercial use
Found your pick? Build a full stack around it — Flowpicker shows compatibility warnings before you commit.
Open the stack planner →