Qwen/Qwen3.8-27B model card

Card read 2026-10-17. States 27B dense causal LM with vision encoder, hidden 5120, 64 layers in the pattern 16 x (3 x Gated DeltaNet + 1 x Gated Attention), Gated Attention 24 Q / 4 KV heads at head dim 256 with RoPE dim 64, Gated DeltaNet 48 V heads and 16 QK heads at dim 128, FFN intermediate 17408, vocab 248320, native context 262144. Benchmark table (vendor, Claude Code harness where noted): Terminal Bench 2.1 73.0 vs Qwen3.6-27B 63.4; SWE-bench Pro 61.7 vs Qwen3.6-27B 53.5; GPQA Diamond 89.2 vs Qwen3.6-27B 87.8. Card says the weights are compatible with Hugging Face Transformers, vLLM, SGLang, and TokenSpeed, and links vLLM and SGLang recipes. HF API createdAt 2026-08-05T08:22:59Z, lastModified 2026-08-14; safetensors total 27,781,427,952 BF16. Benchmark numbers are vendor claims.