What it costs to run Qwen3-Next 80B-A3B (MoE)

How these numbers are produced

• Model shape comes from the model's own config.json — 12 layers, 16 attention heads, 2 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 (10 of 512), so throughput is scored on active parameters, not the full 80B. Memory is the opposite: every expert must still be resident in VRAM, so the figures above use the full weight set.Hybrid MoE: 48 layers, full attention every 4th (full_attention_interval 4) = 12 full-attention + 36 linear-attention layers; layers = the 12 KV layers (linear layers keep a small fixed state). ~3B active params per token, but all ~80B weights stay resident. 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, 10 Oct 2026 08:21:01 GMT). Spot/interruptible pricing can change or vanish without notice.