Running MiniMax M3 on MI355X
Quick answer
MiniMax M3 sustains 29,616 tokens/s per GPU on MI355X at 50 tokens/s per user, which works out to $0.014 per million tokens at hyperscaler pricing, served by vLLM. Fastest measured TTFT: 0.1 ms; fastest TPOT: 0.0 ms (each the best across all configs, not one run).
Benchmarked configs
288
Serving engines
ATOM¹, Mooncake ATOMesh¹, VLLM-DISAGG, vLLM
Precisions
fp4, fp8
Run dates
2026-07-02 → 2026-09-04
Throughput at every interactivity target
Serving is a trade-off: push more concurrent users through a GPU and each user's tokens arrive slower. The ladder below reads the measured frontier at each per-user speed target on the AgentX agentic coding workload, using the best engine and precision at that point.
| Per-user target | Tokens/s per GPU | $ / 1M tokens | Engine | Precision |
|---|---|---|---|---|
| 30 tok/s | - | - | vLLM | fp4 |
| 50 tok/s | 29,616 | $0.014 | vLLM | fp4 |
| 75 tok/s | 25,784 | $0.016 | vLLM | fp4 |
| 100 tok/s | 23,653 | $0.018 | vLLM | fp4 |
| 150 tok/s | 19,119 | $0.022 | vLLM | fp4 |
| 200 tok/s | 9,578 | $0.044 | vLLM | fp4 |
What serving actually costs
Converting the 50 tokens/s per user operating point to $ per million total tokens across rental pricing tiers from the SemiAnalysis AI Cloud TCO model.
| Pricing tier | $ / GPU / hr | $ / 1M tokens |
|---|---|---|
| Owning at Large Hyperscaler Volume | $1.50 | $0.014 |
| Retail | $2.90 | $0.027 |
Frequently asked questions
- How fast is MiniMax M3 on MI355X?
- At an interactivity target of 50 tokens/s per user on the AgentX agentic coding workload, MI355X sustains 29,616 tokens/s per GPU serving MiniMax M3 with vLLM in FP4. Peak measured throughput across all configs is 29,699 tokens/s per GPU.
- How much does it cost to serve MiniMax M3 on MI355X?
- $0.014 per million total tokens at large-hyperscaler-volume ownership $/GPU/hr pricing, at 50 tokens/s per user. The retail rental tier is tabulated above; slower interactivity targets lower the cost further.
- Which serving engines run MiniMax M3 on MI355X?
- The runs behind this page used ATOM¹, Mooncake ATOMesh¹, VLLM-DISAGG, vLLM in FP4, FP8, including disaggregated prefill and multi-node serving. Engines are rebuilt and re-benchmarked continuously, so the best config can change between visits.
- How are these MiniMax M3 numbers measured?
- Every number is measured on real MI355X hardware by the InferenceX fleet, sweeping concurrency on the AgentX agentic coding workload to trace the throughput-versus-interactivity frontier; the newest run landed on 2026-09-04. The same derivation powers the InferenceX overview leaderboard.