MODEL · SCHEMA 6 · AUTHORITY CANONICAL
QWEN3.8-27B
64 language layers that do not all run the same programme, compiled to a precision map rather than a bit-width. Open it.
What is inside Qwen3.8-27B?
Sixty-four language layers, hidden size 5,120, in a hybrid stack: forty-eight layers mix tokens with a Gated DeltaNet recurrence and sixteen use gated attention, in a repeating pattern of three then one. A vision perception tower sits beside the language stack. The container holds five logical objects and declares each layer's operator explicitly, so every address below resolves by role rather than by filename.
STATUS
The strip tells the truth as it changes, and it has. FIDELITY read MEASURING while the bank ran; Q-BANK-1 returned on 2026-08-31 and it reads MEASURED. Vision stays NOT EVALUATED, because the tower's bytes are carried and carried is not measured.
64 language layers, and they do not all run the same programme. Three mix tokens with a Gated DeltaNet recurrence, then one uses gated attention — repeated sixteen times. The label comes from the operator each layer declares in the graph, never from which tensors it happens to ship.
recorded — qwen3.8-27b.s6.vindex3 · encoded and compiled 2026-08-30
The model does not have a precision. It has a precision map.
WHAT WAS COMPILED
The decoder stack compiled from 45.4 GiB to 12.8 GiB — 3.54×, 4.5012 bits per weight derived from the stored bytes rather than asserted. 400 tensors were re-encoded and 448 carried verbatim. With the embedding, norm, output head and vision tower left at BF16, the deployable artifact is 18.4 GiB against 51.0 GiB — 2.77×. Sixteen bits anywhere on that map means the current programme chose sixteen bits, not that anything was proven to need them.
OBJECTS
Five objects, and no multi-token-prediction surface among them. The page names what the artifact holds and nothing else.
SURFACES LEFT AT BF16
EVIDENCE
Representable
Encoded at schema 6 with capability text-generation. Whole-model admission refuses: the vision_config declares neither layer_norm_eps nor rms_norm_eps, so its norm surface is undeclared and the encoder will not invent one. The text stack is fully admissible.
NVFP4 compiled
400 tensors re-encoded in 97 seconds; 4.5012 bits per weight across the decoder stack, derived from stored bytes.
Self-verified
Recorded hashes and the structural record re-derived from the artifact alone, without an inference engine. That is self-integrity — not a proof that the container matches the original checkpoint, which is a separate and stronger claim made at encode time.
NVFP4 execution
The CPU path executes this Gated DeltaNet programme against the NVFP4 representation, so the representation requested is the representation measured — no dequantising to BF16 behind the measurement.
Text-decoder fidelity
Q-BANK-1, 1,740 positions against the BF16 reference: KL mean 0.02411 (median 0.01301, p95 0.08708), ΔNLL mean +0.01725, top-1 agreement 93.91% with 106 flips, top-5 overlap 91.61%.
The recurrence-control exception
Preserving the decay and write projections at BF16 removed 8 top-1 flips of 1,740 — McNemar exact p = 0.0215 on the paired difference — for 33.9 MB, while KL p95 and mean ΔNLL were better without it and the high-margin tertile flipped zero times either way. Real, and too small to earn a default exception. The compiler no longer protects that path.
Vision fidelity
The perception tower's bytes are carried untouched. Preserved bytes are not evaluated behaviour, and the Record says which is which.
WHY THIS PAGE HAS A GAP IN IT
Same top token is not the same model: a representation can choose the identical winning token at every measured position while the distribution underneath it moves. So fidelity here waits on measurements over distributions rather than on whether two generations happen to read alike — and the cell stays empty until they exist.