Best open-weight coding models to run locally

These picks already have weights out. Flint is Recursive's local coding model. Nanbeige4.3-33B is an Apache 2.0 coder. Nemotron 3.5 Bolt is the small model. Beam and Muse Spark Open are large mixture-of-experts models you can self-host. Kimi K3.1 Pro is not ranked: Moonshot says the full weights will be released by 4 Nov 2026. MiniMax M3.1 Flash is not ranked either: MiniMax says the weights will be released later. This page does not add a VRAM number unless the vendor already stated one. The apps that run these weights are on the local AI shortlist.

Related: Best tools to run AI models locally and Ollama vs LM Studio.

Category: Local models and Coding agents.

1. 01FL

   Flint

   Recursive's open-weight mixture-of-experts model for local agentic coding.

   Recursive's open-weight model for local agentic coding. The model card lists about 39.8 billion parameters and about 2.5 billion active, Apache 2.0 weights, and a 262,144-token context. Recursive says it works with Ollama, LM Studio, and llama.cpp.

   FreeCoding agentsLocal models

   4.2/ 5

   Visit website
2. 02N3

   Nanbeige4.3-33B

   Apache 2.0 model for agentic coding and cowork tasks, with a 512K context.

   Apache 2.0 weights aimed at coding and cowork tasks. The model card lists 33,118,977,024 parameters, English and Chinese, and a context window of up to 524,288 tokens, with recipes for Transformers, SGLang, and vLLM.

   FreeCoding agentsLocal models

   4.2/ 5

   Visit website
3. 03N3

   Nemotron 3.5 Bolt E4B

   NVIDIA's open small model for local voice assistants, characters, and device agents.

   NVIDIA's small open model for voice assistants, characters, and device agents, with multi-turn tool use. The research post says target workloads fit in about 4GB. The model card lists 8 billion total parameters, 4 billion effective, a 64k context, and the OpenMDW-1.1 license.

   FreeLocal modelsChat assistants

   4.2/ 5

   Visit website
4. 04BE

   Beam

   Reflection's open-weight model for coding and agents, with a 1M context window.

   Reflection's sparse mixture-of-experts model for coding and agents. The model card lists 501.3 billion total parameters, 23.4 billion active, a 1 million token context, and Apache 2.0 weights. It is the large self-hosted pick.

   FreemiumCoding agentsLocal models

   4.2/ 5

   Visit website
5. 05MS

   Muse Spark Open

   Meta's open-weight multimodal model for coding and agents, with a 1M context window.

   Meta's multimodal model for coding and agents: 816 billion total parameters, 42 billion active, and a 1 million token context. Meta says it runs on two Apple M5 Ultra Mac Studios through MLX. The weights use the Muse Model License, which Meta says has limited commercial terms.

   FreemiumLocal modelsCoding agents

   4.2/ 5

   Visit website

Submit your tool