How much VRAM do you need to run Qwen3-Coder? 24 GB for the 30B, 48 GB for Next, a server for the 480B

How much VRAM do you need for Qwen3-Coder? It depends on which of the three variants you mean. The 30B-A3B is a 19 GB download at Q4_K_M on Ollama and runs fully on a 24 GB GPU at a 32K context. Qwen3-Coder-Next (80B total, 3B active) is a 52 GB download on Ollama and wants a 48 GB card or a Mac with 64 GB or more of unified memory. The 480B-A35B flagship is 290 GB at Q4_K_M, so it's a multi-GPU server model. Sizes read November 3, 2026 on Ollama's library and Qwen's own Hugging Face pages.

How much VRAM does each Qwen3-Coder variant need?

Start from the file size, then add room for context. These are the published file sizes for each variant and quantization, read November 3, 2026. The GPU column is our read of those sizes plus a 32K context; check your own card in the VRAM calculator.

Parameter counts are from each model card on Qwen's Hugging Face page. File sizes are Ollama's tag list, except the 48.4 GB figure, which is the sum of the Q4_K_M files in Qwen's own Qwen3-Coder-Next-GGUF repo.

Is there a Qwen3-Coder 8B or 32B?

No. Qwen's Qwen3-Coder repo lists three model families: Qwen3-Coder-Next, Qwen3-Coder-480B-A35B and Qwen3-Coder-30B-A3B, plus their FP8, GGUF and base builds. The 7B, 14B and 32B coding models are Qwen2.5-Coder, the previous generation. If a guide tells you to run "Qwen3-Coder 32B", it's mixing the two up. An earlier version of this page did too; we corrected it on November 3, 2026.

Does Qwen3-Coder-30B-A3B fit on a 24 GB GPU?

Yes, at 4-bit with a sensible context. The Q4_K_M file is 19 GB. On top of that sits the KV cache, which grows with every token of context. From the model's own config.json (48 layers, 4 KV heads, head size 128), that's about 96 KB per token, or roughly 3 GB per 32K tokens.

Weights and the 16K row are from our cost-to-run page for this model, which adds about 0.8 GB of runtime overhead; the longer rows apply the same KV formula. So a 24 GB card holds the 30B at 32K comfortably and tops out around 64K. For the full 256K window you'd need about 42 GB, or a quantized KV cache. Qwen's own model card says: "If you encounter out-of-memory (OOM) issues, consider reducing the context length to a shorter value, such as 32,768."

The 30B-A3B is a mixture-of-experts model: only 3.3B parameters work on each token, which is why it's quick, but all 30.5B still have to sit in memory. That's why the file is 19 GB even though it "thinks" like a much smaller model. Our best GPU for Qwen3-Coder page compares the 24 GB cards.

Watch out: Ollama gives you 4K or 32K context unless you change it

Ollama picks a default context from your VRAM. Its docs list "< 24 GiB VRAM: 4k context", "24-48 GiB VRAM: 32k context" and ">= 48 GiB VRAM: 256k context", and add: "Tasks which require large context like web search, agents, and coding tools should be set to at least 64000 tokens." A coding agent that seems to forget your files after a few turns is often running at 4K.

How much VRAM does Qwen3-Coder-Next need?

About 48 GB for the 4-bit build. Qwen describes Qwen3-Coder-Next as "an open-weight language model designed specifically for coding agents and local development", with "80B in total and 3B activated" parameters. Ollama's q4_K_M tag is 52 GB and its q8_0 tag is 85 GB; Qwen's own GGUF repo lists 48.4 GB for Q4_K_M and 84.8 GB for Q8_0.

The good news is context. Only 12 of its 48 layers use full attention (the rest are linear-attention layers with a small fixed state), so its KV cache is about a quarter of the 30B's per token: roughly 0.75 GB per 32K tokens, about 6 GB for the full 256K. So once the 45 to 52 GB of weights fit, the full 256K window adds only about 6 GB more. Two 24 GB cards (48 GB total) are tight: our cost-to-run page for Qwen3-Coder-Next puts the Q4_K_M build at 46.1 GB with just 16K of context.

Can you run Qwen3-Coder-480B at home?

Not on a normal desktop. The Q4_K_M build is 290 GB on Ollama, the Q8_0 is 510 GB and the FP16 is 960 GB. Its KV cache is also the biggest: from its config.json (62 layers, 8 KV heads), about 248 KB per token, or close to 8 GB per 32K tokens. At 4-bit and 32K context that's roughly 300 GB of memory, which means several 80 GB data-center GPUs or a 512 GB Mac Studio-class machine. For most people, the 480B is something you call through an API, not something you host.

Can you run Qwen3-Coder on 8 GB, 12 GB or 16 GB of VRAM?

Not fully on the GPU. The smallest Qwen3-Coder is 19 GB at 4-bit, so on 8 to 16 GB you have three honest options:

• Offload the experts to system RAM. Because the 30B-A3B is a mixture-of-experts model, llama.cpp can keep the attention layers on the GPU and put expert weights in system RAM. It runs slower than fully on the GPU, and you need enough RAM for the rest of the 19 GB plus your OS. How it works: --n-cpu-moe explained.
• Drop to a smaller quant. Our cost-to-run page shows Q2_K at about 14 GB with 16K context, which squeezes into 16 GB, but with heavy quality loss. See what quantization does.
• Run Qwen2.5-Coder instead. The 7B and 14B fit 8 to 16 GB cards. See best GPU for Qwen2.5-Coder 14B.

Should you buy a GPU or rent one to try Qwen3-Coder?

Try first if you're unsure. A 24 GB card is the floor for the 30B and a 48 GB card for Next, and both are cheap to rent by the hour. On November 3, 2026 at 20:20 UTC, our price tracker's lowest listings were an RTX 3090 (24 GB) at $0.067/hr and an RTX A6000 (48 GB) at $0.301/hr, both interruptible (spot) offers on Vast.ai; on-demand was $0.096/hr and $0.349/hr. Listings change by the hour; see today's prices on our cloud GPU page.

If you're buying, a used RTX 3090 is the usual cheapest route to 24 GB. For Qwen3-Coder-Next, compare a 48 GB card against a high-memory Mac on the best GPU for Qwen3-Coder-Next page.

Which Qwen3-Coder should you run?

Match it to your memory:

• 24 GB GPU (RTX 3090, 4090) or 32 GB (RTX 5090): Qwen3-Coder-30B-A3B at Q4_K_M, context set to 32K to 64K. A 32 GB card fits Q6_K at 16K context per our cost-to-run page.
• 48 GB GPU or a 64 GB+ Mac (two 24 GB GPUs are tight): Qwen3-Coder-Next at Q4_K_M. Its small KV cache leaves room for long context.
• 8 to 16 GB: the 30B with expert offload to system RAM, or Qwen2.5-Coder 7B/14B.
• 480B: use it through an API; it isn't a home model.

Then set it up in your editor: local AI coding assistant in VS Code or Ollama with OpenCode. Wondering whether it can replace a subscription? See Qwen3-Coder vs GitHub Copilot.