One-line: The open-weight member of the Qwen3.8 generation — a 27B dense hybrid-attention model with a surprise vision encoder, shipped Apache-2.0 on Aug 13-14; head and shoulders above Qwen3.6-27B on agentic coding (Terminal Bench 2.1 73.0 vs 63.4, SWE-bench Pro 61.7 vs 53.5). ⚠ note: pipeline is image-text-to-text (native VL), but it has published coding/agentic evals, so it qualifies for this tracker. All numbers below are lab-reported from the Qwen card (BenchLM independently echoes SWE-bench Pro 61.7 / Terminal-Bench 2.1 73.0).
Active params: 27B dense (64 layers, hidden 5120 — no MoE)
Context length: 262,144 native; extensible to 1M via YaRN scaling
VRAM (fp16 rough): ~54 GB → consumer GPUs need GGUF/int4 quant; the FP8 sibling (Qwen3.8-27B-FP8, 123k downloads) is the practical local pick
Architecture note: Qwen3.5-style hybrid — linear attention interleaved with full attention (1 full per 4 layers) + MTP. Faster long-context inference than pure dense attention. Vision encoder present (native image/video understanding).
What makes it notable
This is the local-runner’s release of the Qwen3.8 generation: 27B dense, Apache-2.0, 262k native context, and the strongest published agentic-coding numbers of any open 27B-class model (SWE-bench Pro 61.7 and Terminal Bench 2.1 73.0 both beat Qwen3.7-Plus, the closed API model). The hybrid attention makes it cheaper to serve at long context than a plain dense 27B. Caveats: numbers are all lab-reported (no independent run yet), it ships as a VL model (vision encoder adds VRAM/complexity for text-only use), and SWE-bench Verified / BFCL are absent — the agentic claim rests on SWE-bench Pro + Terminal Bench 2.1. If you only care about text/coding, the community was already mass-producing GGUF/FP8 builds within 24h of the drop (a strong signal the weights are real and runnable).