Qwen3.8-27B

ADDED 2026-10-17. Dense hybrid in the runnable class (27.8B total, every parameter active). HF API createdAt 2026-08-05T08:22:59Z and card lastModified 2026-08-14, so this is not a two-week release; it is recorded because the catalog had Qwen3.8-Flash-Next and no 27B sibling, and the card's own table beats the previous dense 27B of the line on the benchmarks this job tracks. BEATS Qwen3.6-27B (vendor card, not independently re-run here): Terminal Bench 2.1 73.0 vs 63.4, SWE-bench Pro 61.7 vs 53.5, GPQA Diamond 89.2 vs 87.8. There is no models/qwen3-6-27b record to point at; the comparison is the card's own previous-generation row. SWE-bench Pro and Terminal Bench 2.1 are the card's variants of SWE-bench and Terminal-Bench. AIME is not in the card's table. PARAMS: safetensors total 27,781,427,952 BF16, so params_b and active_params_b are both 27.781. Dense, no MoE keys in config, so active equals total. The card rounds this to 27B. GEOMETRY from config.json text_config, fetched 2026-10-17: hidden 5120, 64 layers, layer_types is 48 linear_attention + 16 full_attention at full_attention_interval 4 (indices 3,7,...,63), num_attention_heads 24, num_key_value_heads 4, head_dim 256 so gqa_ratio 6, intermediate 17408, silu, vocab 248320, tie_word_embeddings false, max_position_embeddings 262144. Linear block: linear_num_key_heads 16, linear_num_value_heads 48, key and value head dim 128, conv kernel 4. RoPE is native mrope: rope_type default, mrope_interleaved true, mrope_section [11, 11, 10], partial_rotary_factor 0.25, rope_theta 10000000, no yarn/ntk factor. context_scaling_method is mrope. KV (bf16, full-attention layers only): 16 x 4 kv heads x 256 x 2 (K and V) x 2 bytes = 65,536 B/token (64 KiB). The 48 linear-attention layers hold a fixed recurrent state, not a growing cache. If all 64 layers cached as GQA this would be 64 x 4 x 256 x 2 x 2 = 262,144 B/token, 4x the recorded figure. flops_per_token_active = 2 x 27.781427952e9 = 5.556e10, the schema one-pass approximation; the linear layers are not 2 FLOPs per parameter, so treat it as the schema convention, not a profile. Vision tower is separate: vision_config depth 27, hidden 1152, patch 16, out_hidden_size 5120. One MTP layer (mtp_num_hidden_layers 1). DEPLOYMENT (card, 2026-10-17): weights are compatible with Hugging Face Transformers, vLLM, SGLang, and TokenSpeed. The card links a vLLM recipe and an SGLang recipe and does not publish a llama.cpp or ollama recipe. No quantization is named on the base card. This repo already has measured decode rows for the checkpoint: ollama's MLX backend on an M3 Max 128GB runs a 4-bit quant at 42.5 tok/s (benchmarks/ollama-m3-max-qwen3-8-27b-mlx-decode-tok-s) and ollama's llama.cpp backend runs Q4_K_M at 12.0 tok/s (benchmarks/ollama-m3-max-qwen3-8-27b-llamacpp-decode-tok-s). So llama.cpp and ollama both run it, at Q4_K_M and at 4-bit MLX respectively. The card itself publishes no tok/s.