Running a capable coding model on your own machine stopped being a compromise sometime in early 2026. Open-weight models now land within striking distance of frontier cloud models on coding benchmarks — and they run on hardware a serious developer can actually afford. No API bills, no data leaving your machine, no rate limits when you are deep in a debugging session.
This guide covers the six open-weight models worth downloading for coding in 2026, what hardware each needs, and how to actually run them with tools like Ollama and LM Studio.
What "open-weight" actually means
"Open-weight" means the model's trained weights are publicly available for download and use — you can run them yourself. It is not always identical to "open source": the training data and code may stay private. What matters for you practically is the license. Apache 2.0 and MIT licensed models can be used commercially without much friction. Meta's Llama community license has restrictions. This guide notes each model's license so you do not get surprised later.
One more practical note: quantization. A model needs roughly 2 GB of VRAM per billion parameters at full precision, but 4-bit quantization (usually Q4_K_M, the default most people run) cuts memory to about a quarter while keeping around 95% of the quality. All hardware figures below assume 4-bit quantized builds unless noted.
1. Qwen3.6-27B — the best all-around local coding model
Scores 77.2% on SWE-bench Verified. License: Apache 2.0. Hardware: single 24 GB GPU (RTX 4090 class).
Qwen3.6-27B from Alibaba is the model most developers should download first. Its dense architecture pairs hybrid Gated DeltaNet with attention, and it famously beats its own 397B-parameter MoE predecessor on SWE-bench Verified — proof that careful architecture beats brute-force size in this range. It also scores 94.1% on AIME 2026 math, handles multimodal vision input, and ships under the clean Apache 2.0 license.
On consumer hardware it fits on a single RTX 4090, and on Apple Silicon it reportedly pushes around 40 tokens per second on an M5 Max. For everyday coding — autocomplete, refactoring, debugging, writing tests — this is the best balance of quality and deployability available offline in 2026.
2. Qwen3.6-35B-A3B — the fast MoE alternative
Scores 73.4% on SWE-bench Verified. License: Apache 2.0. Hardware: 24 GB unified memory (fits on a 24 GB Mac).
If you like Qwen's quality but want faster generation, the 35B-A3B mixture-of-experts variant activates only about 3B parameters per token. That means it generates noticeably quicker than the dense 27B model while keeping a strong 73.4% SWE-bench Verified score. On Apple Silicon, the speed difference matters most on smaller machines where dense-model throughput is the bottleneck.
Running Qwen on Ollama
Both Qwen models are available on Ollama and LM Studio, the two easiest ways to run local models. A typical Ollama setup is two commands:
ollama pull qwen3:32b
ollama run qwen3:32bFor coding-assistant features (tab completion, chat in your editor), pair Ollama with the Continue extension for VS Code or JetBrains. Point Continue at your local Ollama endpoint ("http://localhost:11434") and you get Copilot-style completion with your own model.
3. Devstral Small 24B — the local pick for coding agents
~33 tokens/second estimated. License: Apache 2.0. Hardware: 24 GB GPU.
From Mistral, Devstral Small 24B is purpose-built for agentic coding workflows — models that call tools, run shell commands, and iterate on their own. It is tuned specifically for tool-using agents, making it the closest local stand-in for server-class agentic models. If your workflow is "describe the feature, let the agent work," Devstral Small 2x-class models (the newer generation) are the strongest picks on a single consumer GPU. For building autonomous coding agents on-device, this family is the one to reach for.
4. DeepSeek V4-Flash — the closest thing to a giant, locally
284B parameters (13B active). License: MIT. Hardware: 128 GB Mac (2-bit MLX build is 96.5 GB).
You cannot run DeepSeek V4 Pro locally — it is server-class. But DeepSeek V4-Flash, the distilled 284B-A13B sibling, ships MIT-licensed weights with a 2-bit MLX build of 96.5 GB that fits a 128 GB Mac, reportedly generating around 39 tokens per second on an M5 Max. If you want the DeepSeek coding lineage (the family behind DeepSeek-R1's reasoning models) running offline with serious capability, this is the answer. The MIT license is the most permissive of any major model family here.
DeepSeek also dominates reasoning benchmarks — R1-class reasoning models score around 79.8% on AIME and rival frontier cloud models on logic — so Flash is a strong pick when your coding work leans into hard algorithmic problems.
5. MiniCPM5-2B — the on-device champion
2.52B parameters, 131K context. Scores 69.1 on LiveCodeBench v6. License: Apache 2.0. Hardware: any modern laptop or phone-class device.
Released September 2026, MiniCPM5-2B is the surprise of the year: a 2.52-billion-parameter dense model that scores 69.1 on LiveCodeBench v6 and 46.4 on SWE-bench Verified, with a 131,072-token context window. It runs in llama.cpp and Ollama on ordinary hardware. Do not expect it to architect a system for you — but for inline completion, docstring generation, and quick code questions on a machine with no GPU at all, it is absurdly capable for its size.
6. GLM-5.2 / Qwen3.8-Flash-Next — the server-side open options
Not everything needs to run on your laptop. GLM-5.2 (MIT license) leads long-horizon agentic benchmarks and is the open pick for autonomous coding agents run on your own server. Qwen3.8-Flash-Next offers 262K native context (extendable toward 1M with YaRN) with strong SWE scores — the pick when your work involves very long codebases. If you prefer APIs over self-hosting, DeepSeek V4 Flash is the cost leader at roughly $0.22 per million input tokens off-peak.
How to pick: match the model to your hardware
- No GPU / basic laptop: MiniCPM5-2B in Ollama. Free, tiny, genuinely useful.
- One 24 GB GPU (RTX 4090) or 24 GB Mac: Qwen3.6-27B for the best quality, or Qwen3.6-35B-A3B if you want faster generation. Devstral Small 24B if you run agents.
- 128 GB Mac: DeepSeek V4-Flash via MLX for near-frontier local coding.
- Your own server: GLM-5.2 for agentic work, Qwen3.8-Flash-Next for long context.
Start with one model, wire it into your editor with Continue, and give it a week before switching. The best local coding model is the one actually integrated into your workflow — a slightly weaker model you use constantly beats a stronger one sitting in a downloads folder.