How much VRAM Qwen3 235B-A22B (MoE) needs, and what it costs to run

How these numbers are produced

• Model shape comes from the model's own config.json — 94 layers, 64 attention heads, 4 KV heads.
• Weight size is parameters × effective bits-per-weight. Those constants are checked against real published quantised file sizes — median error 0.7% across 40 measurements.
• KV cache is 2 × layers × KV-heads × head-dim × context × 2 bytes. Using KV-heads rather than attention heads is what makes this correct for grouped-query attention; treating a GQA model as multi-head overstates the cache by up to 8×.
• Tokens/sec is an estimate, not a benchmark. Generation is memory-bandwidth-bound, so this is bandwidth ÷ weight-bytes derated to 75%. Real throughput depends on your runtime, batch size and kernels. We don't run our own hardware tests — see our editorial policy.
• This is a mixture-of-experts model. Generation reads only the experts active for each token (8 of 128), so throughput is scored on ~22.0B active parameters, not the full 235B. Memory is the opposite: every expert must still be resident in VRAM, so the figures above use the full weight set.MoE, standard GQA: all 94 layers hold KV (4 KV heads × 128). 128 experts, 8 per token; active = vendor-published 22B, but all ~235B weights stay resident. Row uses the 2507 Instruct release (262K native context); the April 2025 original has the identical shape but 40,960 native context. HF config.json read 4 Oct 2026.
• Some memory-bandwidth figures are not yet independently verified.Where that's the case we show no tokens/sec at all rather than a number we can't stand behind.
• Rental prices are pulled hourly from provider APIs (last updated Sat, 14 Nov 2026 01:20:08 GMT). Spot/interruptible pricing can change or vanish without notice.