Accelerating Qwen3.6 35B and 27B with GBNF

Qwen3.6 models were compared on an RTX 5090 with Fedora 43, using a llama.cpp mainline build dated April 24th.

• Setup: Qwen3.6-35B-A3B-APEX-I-Balanced.gguf (ctx 216k), Qwen3.6-27B-Uncensored-HauhauCS-Aggressive-Q6_K_P.gguf (ctx 114k), kv f16, -b 256, -ub 256, qwen sampling.
• Tests: a simple "Hi" prompt, a puzzle, and a custom Rust/Next.js 60-task bench.

The gains were larger for 35B-A3B than for 27B.

• 27B: "Hi" tokens dropped from 248→42, puzzle tokens from 40,101→7,376, and puzzle time from 13m36s→2m27s.
• 27B bench score stayed the same at 4620, while bench time dropped from 29m54s→22m20s and throughput improved from 1067→1193 t/s, a +11.8% improvement.
• 35B-A3B: "Hi" tokens dropped from 200→12, puzzle tokens from 30,096→2,592, and puzzle time fell sharply from 2m32s→12s.
• 35B-A3B bench score rose from 4620→4740, a +2.6% increase, and bench time was cut from 33m52s→11m04s, a 67.3% reduction.

In other words, simply adjusting the GBNF grammar can reduce reasoning verbosity in the same model, and for some tasks it can improve both speed and score at the same time.