Qwen3.8-27B scores 29/30 on AIME 2026 with FP8 + xhigh reasoning

A benchmark of the Qwen3.8-27B model on the MathArena/aime_2026 dataset demonstrates that quantized FP8 weights at xhigh reasoning effort achieve a score of 29/30, matching the performance of BF16 models while offering significantly higher throughput.

• Quantized FP8 with xhigh reasoning achieved 29/30 (96.7%), equaling BF16 xhigh but with faster pre-fill (3.4K tk/s) and decode speeds (76 tk/s).
• BF16 medium scored 28/30, while BF16 xhigh also reached 29/30 but required a larger context window of 258,048 tokens.
• The FP8 model outperformed the previous Qwen3.6-27B baseline of 94.1% and matched frontier models like Claude Opus 4.6 and DeepSeek V4 Pro at 96.7%.
• Two problems resulted in empty outputs for both BF16 xhigh and FP8 xhigh due to exhausting the token generation budget rather than producing incorrect answers.