Run Qwen3 VL 235B A22B Thinking locally
Qwen3 VL 235B A22B Thinking is 235B parameters (22B active per token — a mixture-of-experts, so it still needs room for all 235B in memory). Here's the VRAM it needs at each quantization, the smallest GPU that fits, and how a self-hosted box compares with the cheapest verified API.
VRAM by quantization
| Quantization | VRAM needed | Fits on | Notes |
|---|---|---|---|
| int4 (Q4) | 141 GB | 2× H100 (160 GB) ($3.98/hr) | smallest footprint, minor quality loss |
| int8 (Q8) | 282 GB | 4× H100 (320 GB) ($7.96/hr) | near-lossless |
| fp16 (full) | 564 GB | 8× H100 (640 GB) ($15.92/hr) | reference quality |
VRAM ≈ params × bytes/param × 1.2 (overhead). See the fullmethod and break-even calculator.
Self-host vs API
Cheapest API for Qwen3 VL 235B A22B Thinking is $4.4 / 1M tokens(blended) via Alibaba. Running it yourself in int4 fits a2× H100 (160 GB) at $3.98/hr — about $2,905.4/month at 24/7. Those two lines cross at roughly 660M tokens/month: below that the API wins on cost, above it the dedicated GPU does (assuming you keep it busy). Tune your own volume in the break-even calculator.
Run it at home
No cloud account needed: Qwen3 VL 235B A22B Thinking in int4 (141 GB) fits a Mac Studio M5 Ultra (256 GB unified), a $9,499 one-off buy. Amortized over 3 years that's about $0.3615/hr — hardware only, electricity aside. Against the Alibaba API at $4.4/1M, buying it pays for itself after roughly 2.2B tokens total. Below that the API is cheaper; a machine that mostly sits idle rarely earns back its price.
Get the weights
Quantized builds on the model hubs — links search live, so they track new builds as they appear:
- GGUF builds llama.cpp · Ollama · LM Studio
- AWQ builds GPU serving (vLLM)
- GPTQ builds GPU serving
- MLX builds Apple silicon
- Ollama library one-command local run
License: open weights under Apache 2.0 · terms. Compare all providers on the qwen3-vl-235b-a22b-thinking API page.