This article tests the transferability of a Jacobian interpretability lens from Qwen3.6-27B to Qwen3.8-27B, finding that it can read and steer the newer model with zero refitting for specific tasks.
Interpretability lenses get fitted to one exact checkpoint, and as far as I can tell nobody had tested what a version update does to one. So this was my question: when a model line updates, does the fitted instrument survive, or do you refit every release? I tested the published Jacobian lens for Qwen3.6-27B (Neuronpedia, from Anthropic’s July workspace paper) applied unchanged to Qwen3.8-27B. Setup: 3.8-27B shipped 113 days after 3.6-27B. Same 64 layers, same hidden dim, same tokenizer, training relationship undocumented. One protocol, both models, two readouts each: the transported Jacobian readout and the raw logit lens as baseline. bf16, greedy, single seed. Reading result: the main task is 40 two-hop prompts where the middle entity is never stated. Example: “Fact: The currency used in the country shaped like a boot is”, where the target is Italy and Italy appears nowhere in the prompt. The transferred lens keeps the latent entity near the top of the 248,320-token vocab. Median rank at layer 48 is 4 on the home model vs 17 transferred. At layer 24 it’s 121 vs 38, so the successor is actually better at mid-depth (paired sign tests, p < 1e-3). The raw logit lens sits at rank 1e3 to 1e4 through the same band on both models. On WikiText teacher-forced next-token (700 positions), transfer costs 1.2 to 1.3x mid-network and about 2x by layer 48. Latent-content readout transfers nearly clean; surface next-token readout pays more, and pays late. Steering result: I took pullback directions for “ paradox” / “ paradoxical” / 悖论 / 矛盾 from the 3.6 lens, orthogonalized within layer, and projected them out of 3.8’s residual stream at layers 18 to 47 during generation. Prompt: “Describe Escher’s impossible staircase”. The word paradox disappears from the output in all cells, on both models, while the description stays coherent (lithograph, closed loop, illusion all intact). Directions derived entirely from the old checkpoint still find the concept in the new one. Scope: one lens family, one model line, one version step, matched architecture and tokenizer. The design can’t fully separate lens misfit from model change, and I make no claim about cross-family transfer or larger gaps. The practical upshot is that cross-checkpoint transfer is measurable, so a monitoring pipeline can test its lens instead of assuming refit is required. Eval code, the 40-prompt set, per-layer rank tables for all four model-by-readout cells, and the ablation captures: https://huggingface.co/datasets/ec75hash/jacobian-lens-transfer-qwen36-38 Happy to answer questions about the protocol, or hear where you think it breaks.
A hobbyist compares Qwen3.8 and Qwen3.6 AI models in generating ray-tracing code in BASIC, finding that Qwen3.8 iterates to better results independently.
The article evaluates version 3.6 of the Qwen3.8 27B AI model, highlighting regressions in bias/fairness and config ergonomics while noting improvements in spine and capability, concluding it's not a clean upgrade.
Qwen3.8-27B is a new locally-runnable model that reportedly outperforms previous local models and rivals Opus4.6, with benchmark charts shared and links to ModelScope and HuggingFace.
Qwen releases Qwen3.8-27B, an open-weights 27B dense vision-language model with major gains in coding, professional work, and long-horizon agentic tasks, available in FP8 with flexible thinking control.