Best Local Coding Model for RTX 3090 (24GB) in 2026

If you have a 24GB GPU and want to run a coding LLM locally, these are the models that matter in 2026.

Disclosure: Mullama is built by our team (Cognisoc). Editorial policy.

How we picked the winners

For each category, we evaluated quality on standard benchmarks, speed (tokens/sec) on target hardware, VRAM efficiency, ecosystem support, and maturity. All picks are tested hands-on.

The top picks

Decision matrix

Use case	Best pick	Why
Best overall coding	Qwen 2.5-Coder 32B	HumanEval 82%
Best for code agents	Qwen 3-Coder 30B	256K context, agentic
Best on Apple Silicon	GLM-4.7 Flash 30B	Strongest 30B
Best lightweight	DeepSeek Coder V2 16B	16B MoE, fast
Best for agent workflows	Devstral 24B	Mistral, Apache 2.0

Why Qwen 2.5-Coder is still the default at 24GB

Qwen 2.5-Coder 32B sits at the top of this list because a 32B dense model at Q4_K_M is the largest coding model that fits comfortably in 24GB of VRAM with room left for a working context window, and its 82% HumanEval score is the strongest we have recorded for that footprint. Qwen 3-Coder 30B is not simply a smaller substitute — its 256K context window is built for feeding an entire codebase or a long agent trace into the model at once, which the 32B-Coder was never tuned for, so the choice between them is "biggest single-file quality" versus "how much of my repo can the model see."

GLM-4.7 Flash earns the Apple Silicon slot for the same reason the Mac picks elsewhere on this site do: 30B dense on unified memory does not need a 24GB VRAM ceiling to run well, so a Mac Studio with more unified memory than a 3090 has VRAM can use the extra headroom for context instead. DeepSeek Coder V2 16B and Devstral 24B round out the list for narrower jobs — DeepSeek Coder V2 is a 16B MoE built for fast autocomplete latency rather than one-shot quality, and Devstral is Mistral's agent-focused coder, released under Apache 2.0, aimed specifically at multi-step coding-agent workflows rather than inline suggestions.

What a 24GB card cannot do

None of these models are frontier-class next to a hosted GPT-4-tier or Claude-tier coding model on the hardest reasoning-heavy tasks — the gap is real and widest on multi-file refactors that need architectural judgment rather than pattern completion. What a local 24GB setup buys instead is autocomplete latency with no network round-trip, no per-token billing, and a model you can point at proprietary code without it leaving your machine. If your work is mostly large-context agentic refactors rather than autocomplete, budget for the 256K-context Qwen 3-Coder over the higher-benchmark Qwen 2.5-Coder.

Setup notes specific to a 24GB card

All five models here are available as GGUF quantizations, so Ollama is the simplest starting point regardless of which you pick — ollama pull qwen2.5-coder:32b is the whole install step. If you're pairing the model with an IDE, Continue and Tabby both talk to any OpenAI-compatible local endpoint, so switching between these five models later doesn't mean switching editor extensions. Watch context length in practice, not just on paper: a 32B model's stated context window and the context length you can actually run at good speed on 24GB of VRAM are not the same number once the KV cache is accounted for, and pushing close to the stated maximum will slow generation noticeably.

See also

• All learn topics
• All tool reviews
• All comparisons
• All model pages
• Local AI glossary
• All best-of pages