Editorial list
Best open-weight coding models to run locally
Coding models with public weights you can download in 2026. The order below is an editorial pick for local use, not a score. Hardware notes come from the publishers. Where a benchmark number appears, it is lab-reported.
01
Flint
Local pick. Flint is an open-weight mixture-of-experts coding model from Recursive, with about 40 billion total parameters and about 2.5 billion active. The license is Apache 2.0. The MLX INT4 checkpoint is about 22.6 GB. It works with llama.cpp, Ollama, LM Studio, and vLLM. Lab-reported: 74.2% SWE-bench Verified, 54.9% Terminal-Bench 2.1, and 125 or more tokens per second on an M4 Pro MacBook Pro at short contexts.
02
Nanbeige4.3-33B
Nanbeige4.3-33B is a dense 33B model for agentic coding and office tasks, under Apache 2.0, with up to 524,288 tokens (512K) of context. The BF16 checkpoint is about 66.3 GB. Community GGUF quants exist, including a bartowski Q4_K_M of about 20.79 GB. It is dense, so it needs to fit in memory for speed. It runs with Transformers, SGLang, or vLLM. Published benchmarks are lab-reported. This page does not reprint scores. It is the teacher model for a planned smaller Nanbeige4.3-3B. No date is set for that model.
03
MiniMax M3.1 Flash
MiniMax M3.1 Flash weights are public under the minimax-community license, which is not an OSI license. It has about 428 billion total parameters and 23 billion active, so it needs server-class hardware, or you use the API. Context is 1 million tokens.
04
Muse Spark Open
Muse Spark Open weights are public under the Muse Model License, which allows many developers to use the model for free and limits some commercial use. Meta says it can run locally on two Apple M5 Ultra Mac Studios. Context is 1 million tokens.
New and worth a look
Not ranked. Mistral Large 4 is a big-hardware option: about 1T parameters, so most people should use the API or a host. Open weights under Mistral's own license, published October 29, 2026. Mistral's number: 48% on the Artificial Analysis Coding Agent Index. Last checked October 30, 2026.
New, preview
Not ranked. Clover-35B-A3B-Preview is an Apache 2.0 diffusion language model from DiffuSpace, published October 27, 2026. It has 35B total parameters and about 3B active, with a 128K context. BF16 weights are about 69 GB before inference overhead. DiffuSpace's number: SWE-bench Verified 73.6. Last checked October 29, 2026.
Coming soon
Kimi K3.1 Pro is Moonshot's first K3.1 model, described as 3T-class, with a 1 million token context. Moonshot says full weights will be released by November 4, 2026. They are not downloadable yet. It is available on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API.
Harnesses that let you pick a local model
OpenCode
Open source coding agent for the terminal, a desktop app, and an editor. You pick the model.
The app is free and open source (MIT). Model use can still cost money.
Cline
Open source coding agent for editors and the terminal. You choose the model and approve steps.
The agent is open source. You pay the model provider you connect, unless that model is free.
Aider
Terminal pair programmer that maps the repo and commits its edits with git.
Aider is free and open source. You pay for the model API you connect.
Comparison
| Model | License | Size / params | Context | Best for |
|---|---|---|---|---|
| Flint | Apache 2.0 | About 40B total, about 2.5B active. MLX INT4 about 22.6 GB. | Up to 512K. Recursive recommends staying within about 256K. | Local pick for a coding model on your own machine. |
| Nanbeige4.3-33B | Apache 2.0 | Dense 33B. BF16 about 66.3 GB. Community Q4_K_M GGUF about 20.79 GB. | Up to 524,288 tokens (512K). | A dense coding model you can download, if it fits in memory. |
| MiniMax M3.1 Flash | minimax-community (not an OSI license) | About 428B total, 23B active. | 1 million tokens. | Server-class hardware, or the MiniMax API. |
| Muse Spark Open | Muse Model License, with commercial limits | Meta says it runs on two Apple M5 Ultra Mac Studios. | 1 million tokens. | A large open-weight model for coding and agents, if you can run or host it. |
FAQ
- Which open-weight coding model is the local pick?
- Flint. It has about 40 billion total parameters and about 2.5 billion active, under Apache 2.0. The MLX INT4 checkpoint is about 22.6 GB. It works with llama.cpp, Ollama, LM Studio, and vLLM. Lab-reported figures are 74.2% SWE-bench Verified, 54.9% Terminal-Bench 2.1, and 125 or more tokens per second on an M4 Pro at short contexts.
- How large is Nanbeige4.3-33B, and what does a dense model mean for hardware?
- It is a dense 33B model under Apache 2.0, with up to 524,288 tokens (512K) of context. The BF16 checkpoint is about 66.3 GB. A community Q4_K_M GGUF is about 20.79 GB. Because it is dense, it needs to fit in memory for speed.
- Are MiniMax M3.1 Flash and Muse Spark Open free to run anywhere?
- MiniMax M3.1 Flash weights are public under the minimax-community license, which is not an OSI license. It has about 428 billion total parameters and 23 billion active, so it needs server-class hardware, or you use the API. Muse Spark Open uses the Muse Model License, which limits some commercial use. Meta says it runs on two Apple M5 Ultra Mac Studios.
- Can I download Kimi K3.1 Pro?
- Not yet. Moonshot says full weights will be released by November 4, 2026. Kimi K3.1 Pro is a 3T-class model with a 1 million token context. It is available on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API. OpenRouter lists $2 input and $10 output per 1 million tokens. That figure is OpenRouter's listing.