Qwen3.6-27B → Qwen3.8-27B

Qwen3.8-27B is an exact graph-visible continuation of Qwen3.6-27B. The two hosted, trace-backed hfviewer graphs contain the same 59 source-faithful nodes and the same topology. Move through all four granularity levels or switch between split and Unified views: there are deliberately no colored diff annotations because there is no architecture change to annotate.

The architecture stayed fixed

The scores did not

Qwen reports substantially stronger results despite preserving the same graph-visible architecture. The largest gains in this selection are on software engineering and computer use; GPQA Diamond, by contrast, moves only 1.4 points.

Source: Qwen's official Qwen3.8-27B model card, pinned to the release revision, which reports both checkpoints in the same comparison tables. These are Qwen-reported scores, not independent reproductions. QwenSWEBench is an in-house benchmark; Qwen says its SWE-bench Pro baselines were re-evaluated on a refined task set, and WebArena-Verified uses the official grader under the OSWorld scaffold.

What zero means

The official config.json files agree on every architecture-bearing field used by the graph. Their only configuration-file difference is the recorded transformers_version metadata. hfviewer maps every node one-to-one and finds no additions, removals, changed operators, rewired edges, altered repeat counts, or changed graph-visible dimensions.

This is an architecture result, not a claim that the checkpoints are identical. Training data, optimization, learned weights, post-training and behavior can change substantially while the computation graph stays fixed.

How to inspect it

Desktop opens in synchronized split view so the two graphs can be checked directly. Pan or zoom either pane, move from Block through Level 4, then turn on Unified to collapse the identical pair into the Qwen3.8 graph. Mobile uses the compact Unified view while retaining the same four granularity levels.