Top 5 Open-Source Coding Models to Run on Your Mac (2026)

Apple Silicon Macs are the best local-AI hardware most developers already own. Unified memory means a 32GB M-series machine can hold what would require a 24GB NVIDIA GPU, and the Metal and MLX backends in Ollama, LM Studio, and llama.cpp make setup a single command. The question is no longer "can I run a model locally?" — it is "which model?" This guide ranks the five coding models that make sense on a Mac in 2026, with real pull tags, download sizes, and context windows. All figures are verified against the Ollama library as of August 2026.

How to Read the Rankings

Coding quality on open models is measured by two signals: SWE-bench Verified (can the model resolve a real GitHub issue end to end) and Aider polyglot (can it produce correct diffs across languages). HumanEval is saturated and no longer discriminates. On Macs, the second dimension is memory: unified RAM is shared between weights, KV cache, and your browser, so quality-per-gigabyte decides what fits.

1. qwen3-coder:30b — The Default Pick

Spec	Value
Architecture	30B MoE, 3.3B active per token
Download (Q4_K_M)	19GB
Context	256K native (1M via extrapolation)
License	Apache 2.0
Runs on	24-32GB GPU or 32GB Mac

The best quality-per-gigabyte on a Mac. Because only 3.3B parameters activate per token, it has the speed profile of a small model with the quality of a much larger one — on an M3 Max (64GB) under LM Studio, expect comfortably above 20 tokens/sec on code generation. It was trained with long-horizon reinforcement learning on SWE-bench-style agentic tasks, so it shines in the loop coding agents actually run: read a file, search related code, edit across files, run tests. The model card publishes no numeric SWE-bench score; treat unverifiable numbers skeptically, but the agentic behavior is real.

ollama pull qwen3-coder:30b
ollama run qwen3-coder:30b "Find and fix the null pointer in src/auth.ts"

Skip it if: you need a hard, published benchmark number, or you only have 16GB.

2. devstral:24b — The Benchmarked Agent Model

Spec	Value
Architecture	24B dense
Download (Q4_K_M)	14GB
Context	128K
License	Apache 2.0
SWE-bench Verified	46.8%

The only local coder with a hard agentic number on its card: 46.8% SWE-bench Verified, ahead of GPT-4.1-mini (23.6%) and Claude 3.5 Haiku (40.6%). Devstral was co-designed with All Hands AI for agent scaffolds, so it slots into OpenHands, Cline, and Aider cleanly. It is also the lightest of the top three on disk at 14GB, which leaves more unified memory headroom for context. The tradeoff: a 24B dense model is slower per token than the 3.3B-active qwen3-coder MoE at the same RAM.

ollama pull devstral:24b
ollama run devstral:24b "Implement the failing test in tests/test_parser.py"

Skip it if: you want the fastest interactive speed, or you need the largest possible context.

3. gpt-oss:20b — The 16GB Laptop Special

Spec	Value
Architecture	20B MoE, MXFP4 quantized
Download	14GB
Context	128K
License	Apache 2.0
Runs on	16GB of RAM

OpenAI's open-weights MoE family. The 20B ships natively at MXFP4 (4.25 bits per parameter on the MoE weights), which is why 14GB of download runs on a 16GB machine — no discrete GPU required. On a 16GB MacBook Air, this is the best general-purpose model that fits. For dedicated coding on the same hardware, qwen2.5-coder:14b (9GB, 32K context) is the alternative: it trades context for a stronger code-specific prior.

ollama pull gpt-oss:20b
ollama run gpt-oss:20b "Explain this stack trace and propose a fix"

Skip it if: you have 24GB+ and want the code-specific training of qwen3-coder or devstral.

4. qwen3.6:27b — The Newest Agentic Coder

Spec	Value
Architecture	27B
Download (Q4_K_M)	17GB
Context	256K
License	Apache 2.0
Input	Text + image

The newest coding family in the Ollama library (added mid-2026), published as an agentic-coding upgrade over the Qwen3 generation. It ships coding-specific quant tags — qwen3.6:27b-coding-nvfp4 at 20GB and qwen3.6:27b-coding-mxfp8 at 31GB — plus MTP tags that Ollama 0.32.6 uses for automatic speculative decoding on Apple GPUs. A 35B tag (24GB) targets 32GB Macs. No numeric SWE-bench score is published on the Ollama card, so this is a quality-by-reputation pick rather than a benchmarked one.

ollama pull qwen3.6:27b
ollama run qwen3.6:27b "Refactor this legacy class to async/await"

Skip it if: you need a measured benchmark, or prefer the MoE speed profile of qwen3-coder:30b.

5. llama3.3:70b — The Ecosystem Workhorse

Spec	Value
Architecture	70B dense
Download (Q4_K_M)	43GB
Context	128K
License	Llama 3.3 Community License
Runs on	64GB+ Mac

Not a code specialist — on agentic SWE-bench tasks it trails the dedicated coders above — but it is the most widely supported open model in existence. Every runtime supports it, every agent scaffold has been tested against it, and there are quants for nearly every hardware profile. If you want one model that "just works" across RAG, chat, and general coding while you experiment, this is the safe default on a 64GB machine. Expect roughly 5-8 tokens/sec on a 64GB M-series at Q4.

ollama pull llama3.3:70b

Skip it if: your workload is primarily code, or you have less than 64GB.

What About the Others?

• glm-4.7-flash (19GB, 198K context): the newest 30B-class coder in the library; a solid qwen3-coder alternative if you want a dense profile in that class.
• qwen2.5-coder:32b (20GB, 32K context): the prior-generation reference coder; still leads open-source models on multi-language code repair (73.7 Aider repair, 75.2 MdEval) but the 32K context cap shows its age.
• deepseek-r1:32b (20GB, 128K): the reasoning pick for debugging and math-heavy work, but reasoning models emit a thinking trace that adds latency — wrong tool for autocomplete.
• codestral:22b (13GB, 32K): the fill-in-the-middle autocomplete specialist, but ships under the Mistral Non-Production License, so it is a research tool, not a production one.

Picking by Mac Memory

Mac RAM	Best coding model	Best general model
16GB	gpt-oss:20b or qwen2.5-coder:14b	gpt-oss:20b
24GB	qwen3.6:27b / qwen3-coder:30b	gemma4:26b / deepseek-r1:32b
32GB	qwen3-coder:30b / qwen3.6:35b	deepseek-r1:32b
48GB	qwen2.5-coder:32b (q8_0)	llama3.3:70b
64GB+	qwen3-coder:30b (q8_0)	llama3.3:70b

Remember: the download size is the floor. Unified memory also pays for the KV cache — a 70B model at 32K context adds ~14GB on top of weights. If your Mac is also running a browser, an IDE, and Docker, budget 8-16GB of headroom.

Implementation Checklist

• Match the model to Mac RAM including KV cache headroom, not just download size
• Prefer qwen3-coder:30b for interactive agentic coding on 32GB+
• Choose devstral:24b when you need a published SWE-bench number
• Use gpt-oss:20b on 16GB machines without a discrete GPU
• Enable the MLX/MTP speculative decoding path in Ollama 0.32.6+ on Apple Silicon
• For code, prefer q8_0 over Q4_K_M when memory allows
• Cap num_ctx to what the task needs so the KV cache does not steal your model's budget

---

MatterAI builds frontier AI infrastructure for engineering teams — from inference-optimized models to autonomous coding agents and agentic code reviews.

Explore what we're building:

• Orbital IDE — Autonomous AI coding agent with background agents and deep codebase memory
• AI Code Reviews — Agentic pre-commit reviews across GitHub, GitLab, and Bitbucket
• Axon Models — Frontier-grade reasoning models at 70% lower inference cost

Get started free - https://app.matterai.so

• Read the docs
• View pricing

---

Follow us on X · LinkedIn · GitHub