Running Kimi K2.6 on MI355X
Quick answer
Kimi K2.6 sustains 2,831 tokens/s per GPU on MI355X at 50 tokens/s per user, which works out to $0.15 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
105
Serving engines
ATOM¹, Mooncake ATOMesh¹, VLLM-DISAGG, vLLM
Precisions
fp4, int4
Run dates
2026-03-27 → 2026-07-30
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 a single-turn chat workload (8k input / 1k output), using the best engine and precision at that point.
| Per-user target | Tokens/s per GPU | $ / 1M tokens | Engine | Precision |
|---|---|---|---|---|
| 30 tok/s | 3,968 | $0.10 | vLLM | fp4 |
| 50 tok/s | 2,831 | $0.15 | vLLM | fp4 |
| 75 tok/s | 2,016 | $0.21 | vLLM | fp4 |
| 100 tok/s | 1,442 | $0.29 | vLLM | fp4 |
| 150 tok/s | - | - | vLLM | fp4 |
| 200 tok/s | - | - | 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 |
|---|---|---|
| Hyperscaler | $1.50 | $0.15 |
| Neocloud | $2.09 | $0.21 |
| Retail | $2.10 | $0.21 |
Frequently asked questions
- How fast is Kimi K2.6 on MI355X?
- At an interactivity target of 50 tokens/s per user on a single-turn chat workload (8k input / 1k output), MI355X sustains 2,831 tokens/s per GPU serving Kimi K2.6 with vLLM in FP4. Peak measured throughput across all configs is 5,370 tokens/s per GPU.
- How much does it cost to serve Kimi K2.6 on MI355X?
- $0.15 per million total tokens at hyperscaler $/GPU/hr pricing, at 50 tokens/s per user. Neocloud and retail rental tiers are tabulated above; slower interactivity targets lower the cost further.
- Which serving engines run Kimi K2.6 on MI355X?
- The runs behind this page used ATOM¹, Mooncake ATOMesh¹, VLLM-DISAGG, vLLM in FP4, INT4, 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 Kimi K2.6 numbers measured?
- Every number is measured on real MI355X hardware by the InferenceX fleet, sweeping concurrency on a single-turn chat workload (8k input / 1k output) to trace the throughput-versus-interactivity frontier; the newest run landed on 2026-07-30. The same derivation powers the InferenceX overview leaderboard.