Qwen3-Coder 30B-A3B Instruct

Running Qwen3-Coder 30B-A3B Instruct locally

At Q4_K_M and 4K of context, Qwen3-Coder 30B-A3B Instruct needs about 19.7 GB — 17.6 GB of weights, 0.38 GB of KV cache and a 1.8 GB activation buffer. The smallest card in this index that clears that comfortably is the RTX 4090 at 24 GB, and 29 of the 63 cards here do. These are calculated figures, not measurements: the estimate stops counting a card as comfortable at 88% of its VRAM, which is roughly the room a desktop session needs.

What longer context costs

Going from 4K to 32K adds about 2.62 GB, taking the total to 22.6 GB. Weights do not move with context — only the KV cache does, and it grows linearly, so this is the number to watch when planning for long documents. The model's native window is 256K; holding all of it at Q4_K_M would need about 46 GB.

Which build to download

This index tracks 2 formats for it — GGUF, AWQ — across 3 levels. Q5_K_M carries the lowest published perplexity loss at 1.0%. The fastest level measured here is AWQ INT4 at 115 tok/s on an RTX 4090, batch 1. GGUF runs on llama.cpp and Ollama across NVIDIA, AMD and Apple silicon; AWQ and GPTQ target vLLM on CUDA and ROCm; EXL2 is ExLlamaV2 and CUDA only.

Common questions

How much VRAM does Qwen3-Coder 30B-A3B Instruct need?

About 19.7 GB at Q4_K_M with 4K of context and batch 1, rising to roughly 22.6 GB at 32K. That figure is weights plus KV cache plus a 10% activation buffer, calculated from the model's architecture rather than measured on a card.

Will Qwen3-Coder 30B-A3B Instruct run on a 24GB GPU?

Yes — at Q4_K_M and 4K context it needs about 19.7 GB, which leaves 4.3 GB spare on a RTX 4090. That is the smallest card in this index that clears it comfortably; 29 of 63 do.

Which quantization of Qwen3-Coder 30B-A3B Instruct should I use?

Q5_K_M has the lowest published quality loss (1.0%), and Q4_K_M is the level most people run. All 3 levels in the index are Q4_K_M, Q5_K_M, AWQ INT4. Note this is a mixture-of-experts model — all parameters must be resident even though only a fraction are active per token, so the memory cost follows the total, not the active count.