Release Notes
v26.11.1
On this page
Released 2026-11-04 · Release line 26.11 · Commit 0eb83be4b
This release updates the default inference images, fixes inference templates from A100 and V100 runs, and ships third-party license texts in every archive. This release has no database migration.
Upgrade Notes
- The default vLLM runner that inference templates name moves from
cuda13.0-vllm0.30.0-r2to-r4, and the vLLM-Omni runner fromcuda13.0-vllm-omni0.30.0-r3to-r4. New deployments from these templates pull the new images; running deployments are not affected. Offline environments must import the new images first; see Offline Deployment.
Added
- Qwen3-0.6B gains a V100 (Volta) spec on the
cuda12.8-v100-vllm1cat1.5.1-r1image (1Cat-vLLM). Before this, a V100 node had no matching spec and fell back to a generic image without sm_70 kernels. - Each archive ships
LICENSE,NOTICE, andLICENSES/(the license texts of embedded third-party programs such as PRoot, talloc, and perftest) next togpuworksandINSTALL.txt.
Changed
cuda13.0-vllm0.30.0-r4adds decord and moviepy for the ERNIE-4.5-VL remote code;cuda13.0-vllm-omni0.30.0-r4adds voxcpm for VoxCPM2.INSTALL.txtin each archive describes install, the default directory, and upgrade per operating system, drops the removed--public-urlflag, and links to gpuworks.readthedocs.io.
Fixed
- About 20 inference templates get launch arguments and mounts fixed from A100 runs: the MOSS-TTS family, Ministral-3, Mistral-Small-4, Devstral-Small-2, ERNIE-4.5, MiMo-V2-Flash, MiniMax-M2.5, Hunyuan-A13B, Llama 3.x 70B, TranslateGemma-27B, Seed-OSS, and others; the Step-3.5 INT4 spec, whose repository is split GGUF only, is removed.
- GLM-TTS pins the glm_tts pipeline and mounts campplus.onnx; MOSS-TTS-Realtime fixes its weight mount path and its memory split on 24 GiB cards.
- In vLLM-Omni, the MOSS-TTS-Realtime reference encoder handed CPU waveforms to a codec on the GPU, so every request failed with a device mismatch.
Full Changelog
Was this page helpful?