# Nanbeige4.2-3B AWQ gauntlet (gentoo)

- host: gentoo (3× RTX 5060 Ti 16 GB)
- started: 2026-07-22
- gpu: CUDA device 2
- port: 8018
- engine: Nanbeige/vllm `@nanbeige42` (fork commit `01af9909913c38fe9b77fb4be0f019573e4cfb6f`)
- model: [`LostGentoo/Nanbeige4.2-3B-AWQ-W4A16-ASYM`](https://huggingface.co/LostGentoo/Nanbeige4.2-3B-AWQ-W4A16-ASYM)
- base: Nanbeige/Nanbeige4.2-3B (BOSS Zhipin Nanbeige Lab)
- quant: llm-compressor AWQ W4A16_ASYM
- max_model_len: 4096
- gpu_memory_utilization: 0.70 (throughput + quality)
- max_num_seqs: 64–128
- gen length (throughput): 128 tokens
- warmup / repeats: 2 / 2
- context tiers: 128 512 1024 2048 4096

| suite | status | notes |
|---|---|---|
| concurrency sweep | OK | 0 errors, peak ~1442 sum-decode tok/s @ conc=20 |
| context sweep | PARTIAL | 128–2048 OK; 4096 n/a (prompt+128 > max_model_len) |
| quality chat_core | OK | ifeval + gpqa:diamond + mmlu_pro; thinking template defaulted OFF for scoring |


- finished: 2026-07-22T12:14:58