Weeyuga benchmark run: Qwen3.8-27B on 2x R9700
Our contribution to the open archive at benchmarks.weeyuga.com, run with its frozen prompts.
Run 8d63125e-6074-4143-befd-1139add7a797 · 2026-10-20 22:19:51 UTC · 36 calls · 0 errors
What this is
The Weeyuga benchmark archive is run by @WeeyugaFounder. It publishes honest local-AI measurements on hardware people actually own, and anyone can add a run. We ran its three frozen prompts (a one-word “hi”, a 4-sentence explanation and a 200-word story) on our home rig and recorded the same files his runs publish. The prompts, the token caps and temperature 0.1 are his. Each prompt got one cold call followed by 5 warm calls, sent one at a time. If you want to add your own hardware, start at his quickstart.
Hardware and software
- GPUs: 2x AMD Radeon AI PRO R9700 32 GB (rocm-smi: 34.21 GB each), power cap 210 W each (sysfs power1_cap 210000000 uW)
- CPU, memory, board: AMD Ryzen 9 7950X (16C/32T), 64 GB host, ASUS ProArt B850
- Host: Proxmox VE 9.2 (pve-manager 9.2.21), running kernel 7.0.14-24-pve; Debian 13 unprivileged LXC (CT 113, 8 cores, 32 GB RAM limit), vLLM in Docker inside the LXC
- Engine: vLLM 0.27.1, image
stilldeadcode/vllm-radiance:0.9.3, ROCm 7.14 (HIP 7.14.60850), torch 2.11.0+rocm7.14 - Model: Qwen3.8-27B-MXFP4-mtpfp8 (local dir; architectures Qwen3_5ForConditionalGeneration, model_type qwen3_5), AMD Quark MXFP4: fp4 weights and dynamic fp4 activations, per-group 32, e8m0 scales; visual tower excluded; MTP fc in fp8_e4m3
- Draft model: Qwen3.8-27B-DFlash2-FP8 (FP8 e4m3 block-128 of incoai/Qwen3.8-27B-DFlash2), HF commit f02593d0ea2e3b7eedbb831fb26f465b1e6238bb
Results
tok/s = output tokens per second after the first token arrives (the archive's formula). Output tokens include reasoning. TTFT = time to first token. Warm = mean of 5 calls for tok/s; median for TTFT and time.
| Prompt | Thinking | Cold tok/s | Warm tok/s | Cold TTFT | Warm TTFT | Warm time | Hit token cap |
|---|---|---|---|---|---|---|---|
| hello | On | 245.3 | 252.0 | 109 ms | 104 ms | 0.26 s | 0 of 6 |
| P-MEDIUM | On | 166.7 | 179.3 | 105 ms | 98 ms | 2.65 s | 1 of 6 |
| P-HARD | On | 209.6 | 204.2 | 110 ms | 102 ms | 5.12 s | 6 of 6 |
| hello | Off | 244.4 | 247.3 | 97 ms | 94 ms | 0.14 s | 0 of 6 |
| P-MEDIUM | Off | 143.2 | 151.6 | 101 ms | 101 ms | 0.91 s | 0 of 6 |
| P-HARD | Off | 120.7 | 107.5 | 106 ms | 108 ms | 3.44 s | 0 of 6 |
GPU power across each cell, both cards combined (GPU sensor, not wall power): thinking on 374.6 W, thinking off 363.3 W.
How this run differs from the archive's own runs
- Speculative decoding is on. A small draft model proposes 7 tokens at a time and the main model checks them. Speed therefore depends on how predictable the text is. That's why thinking-on comes out faster per token than thinking-off here.
- “Cold” means the first call of the run, not a fresh server start. vLLM was already running and we did not restart it. That's why cold and warm are almost the same.
- Thinking hits the token caps. With thinking on, the story prompt spent its whole 1,024-token budget on reasoning and never wrote the story. The caps are his frozen values, so we kept them.
- Two GPUs, each capped at 210 W. The model is split across both cards (tensor parallel 2), and the server runs in Docker inside a Linux container on a Proxmox host, not on bare metal.
- Our own small runner. It used the archive's prompt file unchanged but its own cold/warm loop, so it could measure time to first token. Requests went through an SSH tunnel, which is included in TTFT.
- Power and hash. Power comes from the GPUs' own sensors, not a wall meter. The 19 GB model file wasn't fully hashed;
metadata.jsonlists fingerprints instead.
The full list is in methodology_deviations_md inside metadata.json.
Raw files
- metadata.json – aggregates, hardware, model and deviations
- run.jsonl – one line per call
- run.log – the run as it happened
- run.md – human-readable summary
Benchmark design, prompts and file format: benchmarks.weeyuga.com by @WeeyugaFounder (public archive). We are not affiliated with Weeyuga.