Qwen3.8-27B scores 29/30 on AIME 2026 with FP8, matching frontier flagships
No_Run8812 · reddit · 2026-08-21
A community benchmark of Qwen3.8-27B on MathArena/aime2026 comparing BF16 vs FP8 weights at medium and xhigh reasoning effort:
- FP8 xhigh hits 29/30 (96.7%), tying BF16 xhigh while decoding jumps from 28 to 76 tk/s with much faster pre-fill;
- FP8 medium scored 26/30 vs BF16 medium's 28/30 — slight quantization cost at medium effort;
- On problem 7, both xhigh runs exhausted the token budget with no final answer (empty, not wrong).
Against frontier models (single pass@1): GPT-5.6 Sol xhigh reported 99.9%, GLM-5.2 and GPT-5.4 99.2%, Gemini 3.1 Pro 98.3%, Claude Opus 4.6 and DeepSeek V4 Pro 96.7% — a 27B FP8-quantized model matching several flagships.
Settings: exact-match scoring, temperature zero, sampling disabled, identical chat template and prompt format.
More from Models
- Model at capacity for a straight week drives frustrated dev back to Opus — joonasvirtanen · 2026-10-16
- Decision models match LLMs on routing and safety — and are 7-8x faster — randal_olson · 2026-10-16
- Xiaomi MiMo V2.6 Pro takes #1 open-model spot on Agent Arena at $0.11/task — arena · 2026-10-16
- ChatGPT's 'best AI video tool' answer points to Runway integration, sparking ad doubts — tlakomy · 2026-10-16
- Microsoft's MAI finishes in-house 3T-parameter pretrained base models, post-training results coming — soumitrashukla9 · 2026-10-16
- Hands-on with Gemini 4 Argon: faster than Opus 5.5 but hallucinates more and costs more — prasenx · 2026-10-16