CI / Classify changes (push) Successful in 12s
CI / Web (lint + typecheck + i18n + build) (push) Successful in 1m43s
CI / Format (push) Successful in 8s
CI / Clippy (push) Successful in 2m26s
CI / Test (push) Successful in 13m5s
CI / CUDA type-check (push) Successful in 29m5s
CI / Build cortex SRPM (push) Has been skipped
CI / Publish cortex to COPR (push) Has been skipped
CI / Build neuron SRPM (push) Has been skipped
CI / Publish neuron to COPR (push) Has been skipped
CI / Bump version in source (push) Has been skipped
Activation on beast now loads Qwen3.8-27B rather than the 3.6-27B, and the helexa/large tier points at it (models.toml, deployed separately — it is operator config and not tracked here). The load spec is unchanged because the models are architecturally the same: identical model_type, layer counts, attention layout, vision tower and KV arithmetic, with the two configs differing only in `transformers_version`. Same q6k, same tensor_parallel 2, same devices. Validated on the fleet before switching: loads in ~1m38s first attempt with no code changes; text, streaming and vision all correct, including cell-accurate spatial grounding on a grid image; tool calling end to end through a full round-trip, not just an emitted call; and reachable through the public chain rather than only the mesh. The 3.6-27B stays in the catalogue at the same residency class rather than being removed. Equal rank is what lets the two displace each other in both directions, so a request for either takes beast and the other can take it back — verified in both directions at ~1m20s. Reverting is therefore a one-line alias change, not a redeployment, which is the point of having done the residency work first. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0165r11RzqkMqWWXfJE8tAVU