Kimi-K3.1-Pro on Paiton, measured with LocalMaxxing
6x AMD Instinct MI355X, every GPU capped at 800 W (rated 1,400 W). 11 November 2026.
Not published on localmaxxing.com. We ran LocalMaxxing's open benchmark CLI
(lmx 0.1.60, llm-inference-bench 0.7.7) against Paiton's OpenAI-compatible endpoint. Their submission API only
accepts GPUs from a fixed list, and for AMD Instinct that list stops at the MI300X and MI325X, so the MI355X is
refused. Paiton is also not on their engine list and would go in as "custom". The run files are saved and will be
submitted once the MI355X is added.
Setup
- GPUs
- 6x AMD Instinct MI355X, 288 GB each, 800 W cap per GPU; 2 of the node's 8 GPUs unused
- CPUs, RAM
- 2x AMD EPYC 9575F, 2.2 TB
- Model
- moonshotai/Kimi-K3.1-Pro, released MXFP4 checkpoint
- Engine
- Paiton (build b94aef4f): our own HIP kernels, DSpark speculative decoding, 1M-token context, up to 16 sequences
- Bench
- lmx speed-test (canonical prompt reasoning-v1) and lmx speed-test matrix (30 s per decode cell, max 2,048 output tokens, prefix cache busted per request)
Canonical prompt runs
183 tok/sone request, output
116 msone request, time to first token
628 tok/s16 requests together, 256 output tokens each
Matrix run
Prefill (one cold request)
| context | prompt tokens | TTFT (s) | prefill tok/s |
|---|---|---|---|
| 8K | 8,198 | 0.64 | 12,844 |
| 32K | 32,028 | 2.40 | 13,357 |
| 64K | 63,806 | 4.76 | 13,392 |
| 128K | 127,359 | 9.53 | 13,370 |
Decode throughput, all requests together (output tok/s)
30 s of sustained decode per cell, counted from the server's streamed usage.
| context \ concurrent | 1 | 2 | 4 | 8 | 16 |
|---|---|---|---|---|---|
| 0 | 115 | 176 | 324 | 462 | 741 |
| 8K | 90 | 131 | 241 | 447 | 711 |
| 32K | 82 | 130 | 223 | 381 | 653 |
| 64K | 83 | 125 | 231 | 422 | 663 |
Decode speed seen by each request (tok/s, median)
1 / inter-token latency.
| context \ concurrent | 1 | 2 | 4 | 8 | 16 |
|---|---|---|---|---|---|
| 0 | 130 | 90 | 83 | 57 | 46 |
| 8K | 93 | 67 | 60 | 53 | 39 |
| 32K | 90 | 70 | 56 | 40 | 28 |
| 64K | 89 | 72 | 52 | 43 | 28 |
Time to first token (s, median)
Includes waiting for other requests' prefill.
| context \ concurrent | 1 | 2 | 4 | 8 | 16 |
|---|---|---|---|---|---|
| 0 | 0.09 | 0.11 | 0.11 | 0.49 | 0.96 |
| 8K | 0.63 | 0.65 | 0.96 | 2.87 | 5.40 |
| 32K | 2.44 | 2.45 | 4.84 | 11.00 | 20.82 |
| 64K | 2.41 | 2.42 | 5.95 | 10.80 | 20.44 |
Source: the LocalMaxxing results table supplied by ElioVP for this article. The original run JSON files are not included in this web page.