Qwen 3.8 27B community quants benchmarked on RTX 6000 vs Claude Opus 4.6
Top-Eye-8104 · reddit · 2026-08-27
A co-founder of atomic.chat rented three RTX Pro 6000 96GB cards and ran three community quants of Qwen 3.8 27B (atomic ad-q6k, unsloth ud-q6kl, bartowski q6k), giving each 4 identical prompts — self-playing 3D pool, air hockey, foosball, bowling scoring — asking for a single self-playing HTML game, no system prompt, reasoning at xhigh, all with a dflash2 drafter, best attempt kept.
Results:
| Quant | Size | Total tokens | Avg t/s |
|---|---|---|---|
| atomic ad-q6k | 23.29 GiB | 393,089 | 114.17 |
| unsloth ud-q6kl | 22.53 GiB | 363,083 | 70.33 |
| bartowski q6k | 21.85 GiB | 325,700 | 79.71 |
| Claude Opus 4.6 (subscription) | — | 200,565 | 72.47 |
Atomic's dynamic quant leads clearly on throughput (114 t/s), even above subscription Claude. Prompts and logs are open-sourced on GitHub.
More from Models
- Minimax H3 Max Model Now Available on fal Platform — OdinLovis · 2026-08-27
- tszzl on the HF incident: models metagame tactically but lack strategic awareness — morqon · 2026-08-27
- Anthropic reportedly releasing Fable 5.1, claimed 3-4 months ahead — bindureddy · 2026-08-27
- Google's new Gemma models breeze through Google's own reCAPTCHA v2 — Hour-Wish8158 · 2026-08-27
- Zhipu GLM-5.3-Flash launches on OpenRouter with 1M-token context — AccBalanced · 2026-08-27
- GLM 5.3 Flash now matches Sol 5.6 (Max) on Artificial Analysis' Agentic Index — PilgrimofHaqq2 · 2026-08-27