Qwen3.8-27B Text Record

Exact BF16 hybrid execution and finite-candidate readout, with text-only scope.

Up

Source and compiled representation

Qwen's current text target is Qwen/Qwen3.8-27B, revision 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0, interpreted by the qwen3_5 model owner and graph recipe. The 64 text layers compose 48 recurrent sequence mixers and 16 full-attention layers. Text roles consume 851 of 1,199 source tensors; 333 vision and 15 MTP tensors do not enter the admitted text artifact. Source presence does not publish input capability. Adapter tests and architecture tests guard that boundary.

The unchanged 53,815,809,152-byte BF16 GGUF artifact 1fce07008eaa78e04eedd1a031144f48eb6af617f2b5c508811ba91dca7e00f1 is associated with immutable published release yailabs/Qwen3.8-27B-Text-GGUF@066eb288bffd5a07c0d5ca584114a1f3fcfd13a8 and a fresh authenticated binding v17. Ordinary target-only CUDA execution consumes the 1,732-step compiler-owned forward and two-step output programs, produces 248,320 finite logits, and captures/attaches common prefix schema v2 with all 16 attention and 48 recurrent layers represented. Malformed retained v16 canonical records remain refused. The ordinary admission control preceded the separate readout breadth qualification below. It does not establish upstream whole-model conformance.

Readout and qualification

The completed Qwen breadth Task uses the same common readout schema and score owner as the Mamba control. One captured prefix includes attention and recurrent state; isolated candidate sessions teacher-force exact tokenizer-bound tokens. Sampling/generated counts are zero and one backbone remains resident.

Independent long-double likelihood arithmetic, full-prefix replay, candidate order, cancellation/retry and unchanged source-state controls qualify this exact scope. Retained observations own the numerical observations; allocation projections retain bounded state measurements. These are not calibrated confidence, upstream whole-model conformance or a performance advantage.

Shared owners and limits

Compiler context owns the nested physical program. Common runtime owners derive and re-admit profile identity, capacity, hybrid state, resource projection and cleanup. Qwen does not gain a private scoring runtime, prefix type or scheduler.

Only the exact admitted text representation is covered. Vision/MTP tensors in upstream source do not confer input/output capability. Broader models, public readout APIs, calibrated results, behavior evaluation and release qualification remain separate obligations.