Qwen3.8-27B Q4 on an 8GB RTX 4060 scores 29/30 on AIME 2026 with three no-fine-tune loops
Ai_Glitch_3301 · reddit · 2026-11-19
Setup
Two weeks testing stock Qwen3.8-27B (UD-Q4KM GGUF, llama.cpp) on an 8GB RTX 4060 — no fine-tuning, pass@1, one seed, misses published. MTP draft head boosts speed 3.5→5.7 tok/s (+65%); only 1-in-4 layers keep KV, passing 60k-token recall within 8GB.
Plain model (thinking on)
- AIME 2026: 29/30 (96.7%), 76 min/problem
- HMMT Feb 2026: 28/33 (84.8%); all misses ran out of thinking budget, 2 recovered at 48k (not counted)
- IFEval greedy: 80.2% prompt-strict
- AIME took 38 GPU-hours on a gaming PC overnight
Write-run-fix loops (thinking off, weights untouched)
- Math: model writes/runs Python, 2 crash-fix tries — 136-problem locked set 9 errors → 0, 183s → 22s per problem; harder custom set 7/40 → 36/40
- Code: model runs docstring examples against its own code — HumanEval 90 tasks: 2 → 1 errors; some HumanEval examples are wrong, fixed by letting the model reply KEEP after hand-verifying
- IFEval: model writes its own check(response) and rewrites on failure — locked set 77.4% → 87.0% (27 fixed, 1 broken)
Honest caveats
- The math loop's rival is the thinking-off model; thinking-on matches it, the loop's win is speed (4.6×)
- Coding win was 1 task (noise-level); 96/136 math problems came from the author's earlier benchmark, only 40 truly fresh
- Sandbox is a blocklist + 10s timeout, not a container
Code, prompts, logs: github.com/aiglitchsouls/loop27b
More from Infra
- FT: Anthropic needs ~$602B yearly revenue to earn 20% on $1T compute commitments — rohanpaul_ai · 2026-11-19
- FT: Anthropic needs ~$602B yearly revenue to earn 20% on ~$1T compute commitments — rohanpaul_ai · 2026-11-19
- Linux Mint website goes dark, showing hosting placeholder pages — kingslyj · 2026-11-19
- Tencent Hunyuan's Hy4-Flash engine PR reveals tensor parallel, MTP drafts, CPU expert lane — TencentHunyuan · 2026-11-19
- Commentator calls Jensen Huang's China compute claim bullshit: won't outpace rest of world — teortaxesTex · 2026-11-19
- TRELLIS.2 CUDA kernels on HF Kernels cut BVH build from 7.1s to 5ms — ariG23498 · 2026-11-19