Run Qwen3-Coder Locally: 8B, 80B-A3B, and the 480B Question

Written by Jakub Rusinowski · Last updated 2026-07-12 · Hardware figures computed by our VRAM engine

Qwen3-Coder 8B runs on any 8 GB GPU (`ollama pull qwen3-coder:8b`). The 80B-A3B — the best coder most people can self-host — wants ~49 GB all-in-VRAM (2×24 GB or a 96 GB+ Mac) but runs on ~8 GB VRAM + 32 GB RAM via llama.cpp expert offload. The 480B-A35B flagship needs ~290 GB: admire it over an API. All three are Apache 2.0.

Qwen3-Coder is the family this hub recommends more than any other, for one reason: it's the strongest *dedicated* open coding line you can actually download — the 480B flagship set the open state of the art on SWE-bench Verified (69.6%), and the smaller sizes inherit its training. Apache 2.0 across the board, fill-in-the-middle support for autocomplete, and first-class Ollama tags.

The family spans a 36× size range, so "can I run it?" has three very different answers:

The three sizes, honestly

Q4_K_M figures from our compute engine. The 80B-A3B row shows the all-in-VRAM (fastest) case — see the offload section below.

Getting started in two commands

With Ollama installed:

ollama pull qwen3-coder:8b
ollama run qwen3-coder:8b "refactor this to use a context manager: ..."

For the 80B on capable hardware:

ollama pull qwen3-coder:80b-a3b-q4

Then wire it into VS Code with Continue or Cline — the end-to-end setup guide covers the config. For agent work, raise the context window (16K minimum) as described there; Qwen3-Coder supports long contexts, but your KV-cache VRAM is the real ceiling — model your exact card + context in the checker.

The expert-offload trick: 80B-A3B on a gaming PC

Because only ~3B parameters activate per token, llama.cpp (and recent Ollama builds) can keep the shared layers plus active experts in VRAM and page the rest from system RAM. In practice: ~8 GB VRAM + 32 GB system RAM runs the 80B-A3B at usable-but-not-snappy speeds — a couple of tokens per second slower than all-in-VRAM, dramatically better quality than anything else an 8 GB card can touch.

When it's worth it: delegated tasks where you fire an Aider/Cline job and review the result. When it isn't: interactive autocomplete — keep the 8B for that. This is the single best quality-per-dollar trick in local coding right now, and it's why we list system RAM as a real spec, not an afterthought.

Which size should you run?

• 8 GB card / 16 GB Mac → 8B as your everything-model. Add StarCoder2 3B only if you want even faster ghost text.
• 16–24 GB card → 8B for autocomplete, and honestly consider Qwen 3.6 27B or Devstral-2 22B for chat/agent work — at this tier the general Qwen 3.6 line edges the mid Coder sizes.
• 32 GB+ RAM regardless of GPU → try the 80B-A3B with offload for delegated tasks. It costs nothing but disk space to test.
• 2×24 GB / 96 GB+ Mac → 80B-A3B all-in-VRAM. This is the best local coding experience currently attainable without a server room.
• Need the 480B → hosted endpoints (see the cloud directory) — same weights, same Apache 2.0, someone else's 290 GB.

Hardware that fits the family

The 8B is happy on almost anything; the 80B-A3B all-in-VRAM is the reason 24 GB cards get bought in pairs. If you’re specifying a machine around this model line:

No hardware? Rent the GPU first

Want to feel the 80B-A3B before committing to hardware? Rent a 48 GB+ instance for an hour and run it all-in-VRAM.

• RunPod — RTX 4090 ≈ $0.3–0.7/hr, A100/H100 by the hour; serverless per-second billing available
• Vast.ai — Marketplace of hosted GPUs — often the cheapest per hour (interruptible options)
• Lambda — On-demand A100/H100/B200 instances and clusters, per-hour billing

Full list on the cloud AI directory.

Frequently asked questions

Keep going

• Qwen3-Coder — full spec sheet, quants, and benchmarks
• Report #2 — where Qwen3-Coder sits in the open landscape
• MoE models explained — why 3B active parameters is the trick
• Run a Local Coding Agent End-to-End: Ollama + Qwen3-Coder + VS Code
• Best Local Coding LLMs in 2026, by VRAM Tier
• Local AI coding assistant — the full hub