Yunshu is a local LLM/VLM inference engine that runs on top of MLX for Apple Silicon, and it exposes OpenAI- and Anthropic-compatible APIs. The commit dated November 1, 2026 (3aa49cd) touches only CI configuration, not runtime code. In `.github/workflows/ci.yml`, a single line in the CPU-only console test dependency installation step was changed so that the packages pulled in by the OpenAPI dump are installed at pinned versions, without omission.
This is the step where, because MLX is unavailable on Linux CI, the console is built with a CPU-only Python in order to dump the OpenAPI schema. Alongside the existing fastapi 0.143.0, httpx 0.28.1, uvicorn 0.54.0, websockets 17.2, numpy 2.5.4, python-multipart 0.0.32, xxhash 4.0.1, jsonschema 4.26.0, openai 3.31.0, psutil 7.2.2, regex 2026.9.29, lxml 6.1.3, prometheus-client 0.26.0 were added. The commit message states that this list was verified by actually installing it in a new virtual environment. This Python is also registered as the documentation-generation variable (DOCS_PYTHON), so the change is one that surfaces missing imports in the OpenAPI dump at the CI stage first.
That said, this commit is a one-line edit to the dependency list and does not touch the inference path. The lossless speculative decoding based on DFlash2/MTP, or the performance of the two-tier RAM·SSD prefix cache (hybrid-model prefix cache) described in the repository description, cannot be confirmed by this change, and no benchmark numbers are presented either. The CI console build still runs on a CPU-only Python, so the MLX path remains intact. The license and popularity figures are also absent from the material. The commit message additionally carries the notation 'Co-Authored-By: Claude Opus 5.5'. This change is closer to environment tidying within the range where schema generation runs reliably, and it is too early to read it as an improvement directly tied to inference performance.