Release Notes
v26.11.0
On this page
Released 2026-11-03 · Release line 26.11 · Commit 2c794db9b
The first version of the 26.11 release line and the first GPUWorks release. Earlier builds were published only on the nightly channel.
Main Capabilities
- One static binary each for the server and the agent: the server embeds the web console (Chinese / English) and the agent for each platform, and keeps its state in one SQLite file; agents connect outbound, and the server can dial nodes whose outbound access is restricted (server_dial).
- Heterogeneous accelerators in one scheduler: NVIDIA, Huawei Ascend, T-Head PPU, Hygon DCU, and Apple Silicon nodes join one cluster; the scheduler estimates the memory for weights, KV cache, and engine overhead before choosing GPUs, and explains the placement.
- Inference template catalog: 390 models and 1,224 launch specs (NVIDIA 775, PPU 196, Ascend 173, Hygon 42, Apple 9), each pinning the image, engine, launch arguments, and weight revision.
- Several models per GPU: dGPU hard memory limits (NVIDIA on HAMi-core, Ascend on vcann-rt, PPU on ppu-slicing-shim, Hygon experimental), and vLLM sleep/wake.
- Scale-out: replicas across nodes, multi-node tensor / pipeline parallelism, and PD disaggregation.
- OpenAI-compatible gateway: chat, completions, embeddings, rerank, audio, image, video, and realtime under
/v1, with API tokens and per-token usage. - Hybrid public and private cloud: sync rented machines from RunPod, CompShare, AutoDL, Vast.ai, and Alibaba Cloud, or virtual machines from ZStack Cloud.
- Observability and benchmarking: GPU and engine metrics, log search, per-request tracing (exportable over OTLP), a web terminal, built-in benchmarks, an AI assistant, and a read-only MCP endpoint.
- Works offline: import images and models from archives or HTTP file servers, and sync them between nodes.
- Signed multi-source distribution: the release index carries an ed25519 signature, and install and upgrade choose an available source among
dl.gpuworks.ai, Alibaba Cloud OSS, and Huawei Cloud OBS.
Getting Started
Was this page helpful?