hongfaqiu/strata-manager released v0.2.41. Strata Manager is an open-source desktop app for Windows/Linux x64 that manages the models and services of the Niko1221/Strata local LLM inference engine, checks hardware and API usage, and provides chat and inference tuning. Commit 6bbb0791 is recorded on 2026-11-05, with 20 lines added and 14 lines removed across 10 files.
The functional axis of this release is quantization support matched to inference engine 0.2.11. Quantization is the method of replacing weights with numbers of lower precision in order to reduce memory usage, and the list now includes the seven sets of Unsloth Dynamic v3.1: UD-IQ1_M, UD-Q2_0, UD-IQ2_XS, UD-Q2_K_XL, UD-IQ3_XXS, UD-Q3_K_XL, and UD-IQ4_XS-v31, which requires engine 0.1.57 or later. UD-IQ4_XS-v31 is handled separately from the existing three-part UD-IQ4_XS, and the hub filename is still UD-IQ4_XS, so it is distinguished by whether it is four parts. If the last part of the same group — the roughly 28.8 GB n-gram table — has already been received, it is reused based on content, and when it is received a hard link and a .done file are left behind so that the disk is not occupied redundantly.
There are two adjustments on the Qwen3.6 side. The number of parallel slots is no longer recomputed with the per-slot VRAM value of Flash-Next, and instead uses own_path_parallel_slot_gb as calculated upstream. On a 128K context basis it is about 5.56 GiB per slot, based on upstream PR #5239. Also, the Qwen3.6 architecture currently ignores the --vram-reserve-mib argument, but if upstream #5216 is merged the argument will be rejected, so this option is no longer passed, and it is no longer indicated in the tuning screen or in one-click acceleration.
That said, this commit itself merely raises the version strings in Cargo.toml, Cargo.lock, package.json, and tauri.conf.json from 0.2.40 to 0.2.41, and moves changes that had not yet been released into the 2026-11-05 entry of the CHANGELOG. No performance measurements are presented, and the screenshot manifest states that actual model execution and GPU benchmarks, and native window acceptance checks, are out of scope. Since the slot calculation relies on the upstream formula and the option removal presupposes an upstream merge, it may change again when the engine is updated. The content of the 0.2.41 entry is Chinese only, so there is no English release note in this commit.