Model components / Inference code

Chitu

Chitu is a production-grade inference engine from Tsinghua's PACMAN group focused on running large models across many kinds of hardware, from pure CPU through single GPU to multi-node clusters with prefill-decode separation. It serves DeepSeek, Qwen, GLM and Kimi models on NVIDIA GPUs and on Chinese accelerators including Ascend 910B, Muxi, Hygon and Moore Threads, with FP8 and FP4 weights converted online to run on hardware without native support. A single chitu.run executable launches multi-node and multi-instance deployments.

Default branch is public-main. README is Chinese-primary with an English copy under docs/en/. Verified 2026-09-01 via GitHub, the LICENSE body and the repository README.

Openness

5 high confidence

5.0

license: Apache-2.0(OSI
source: public
core-gated: ungated(paid professional services offered by email are support beside the published engine, not a withheld component

The LICENSE body is the full Apache-2.0 text with only the appendix copyright slot filled ("Copyright Qingcheng.AI") - a one-line diff against the canonical text, no appended condition. The repository is public and active and is the engine itself; the README's commercial line offers professional technical service by email, which is selling support, not gating a core. Vendored third-party code is declared SPDX-style under LICENSES/ and third_party/, the REUSE-adjacent shape.

https://raw.githubusercontent.com/thu-pacman/chitu/public-main/LICENSE recorded 2026-09-01
LICENSE body, read in full and diffed against the canonical Apache-2.0 text: identical except the appendix copyright line reads "Copyright Qingcheng.AI". No appended terms.
https://api.github.com/repos/thu-pacman/chitu recorded 2026-09-01
Repo metadata for thu-pacman/chitu - archived false, license Apache-2.0, 2,997 stars, default branch public-main, pushed 2026-09-01.
https://raw.githubusercontent.com/thu-pacman/chitu/public-main/README.md recorded 2026-09-01
README (Chinese), read for a paid tier (professional services by email only - no gated build): "生产级大模型推理引擎" (production-grade LLM inference engine); milestones - DeepSeek-R1 671B single-GPU via CPU+GPU heterogeneous inference, FP4-to-FP8/BF16 online conversion, Ascend 910B native support, Moore Threads adaptation, chitu.run single-file multi-node launcher; licence section states Apache-2.0 with SPDX-marked third-party snippets under LICENSES/ and submodules under third_party/.

Adoption

2 low confidence

2.0

2,997 GitHub stars, in the 1K-10K band of the stars scale. Distribution is the repo and its release executables, with no package registry channel to count. Acknowledged production use in Chinese enterprise deployments (China Telecom and others credited in the README) is standing the star count may understate; raising the level would need its own evidenced judgment.

https://api.github.com/repos/thu-pacman/chitu recorded 2026-09-01
Repo metadata - stargazers_count = 2,997 - for thu-pacman/chitu.

Capability

4 medium confidence

4.0

One band below the vllm anchor, level with xllm, rtp-llm and fastdeploy, the other production engines whose distinguishing surface is domestic-accelerator breadth. Its hardware span (five vendors plus pure CPU) and PD-separated cluster serving clear the two_below locals; it is not the community-wide throughput frontier the anchor and sglang occupy.

https://raw.githubusercontent.com/thu-pacman/chitu/public-main/README.md recorded 2026-09-01
Milestones and positioning: v0.6.0 chitu.run single-file multi-node/multi-instance/PD- separation launcher; v0.5.x cluster performance and Moore Threads support; v0.4.0 Ascend/NVIDIA/Muxi/Hygon adaptation serving DeepSeek/Qwen/GLM/Kimi; v0.2.2 single-GPU DeepSeek-R1 671B via CPU+GPU heterogeneous inference; v0.3.0 FP4-to-FP8/BF16 kernels.

Verified 2026-09-01