athrael-soju/narwhal released version 0.8.0. This repository is a Python project that handles LLM inference in a role-free fleet — that is, without pre-assigning a role to each slot — by splitting inference into prefill (precomputation) and decode (token generation). The release commit d92e809 is recorded at 2026-10-27T21:21:57Z, while the date on the 0.8.0 entry left in the CHANGELOG is 2026-10-20.
Three features were added in this version. Operator visibility and setup (#384) was added, along with the ability to reuse still-valid calibration values and profiles as they are (#373). Then it also includes a fix that re-secures prefill capacity during a load surge and resolves decode pressure and slot placement issues (#364).
On the packaging side, it was fixed so that characterization golden records and the documentation workflow are both included in the sdist distributions (#371, #369), and the per-step test time budget was increased as well. The version strings in .release-please-manifest.json, CITATION.cff, and pyproject.toml were all raised from 0.7.0 to 0.8.0. The distribution name is narwhal-inference and the license is Apache-2.0.
However, this commit itself consists solely of the version string updates and the added CHANGELOG entry, and the changes amount to no more than 18 insertions and 3 deletions. Since no performance figures or benchmark results are included, it cannot be known from this commit alone at what scale 0.8.0 is advantageous or by how much. Whether the operator screens and setup actually match the real operational flow will need to be confirmed by installing the distribution.