glm-4.5

GLM-4.5, GLM-4.6 and GLM-4.7 form an open-source family of foundation models from Zhipu AI (zai-org) aimed at agentic, reasoning and coding workloads. The repository documents model variants, download links, hardware requirements, deployment commands and evaluation guidance. Model family - GLM-4.5: 355B total parameters with 32B active (MoE), hybrid reasoning with thinking and non-thinking modes. - GLM-4.5-Air: 106B total, 12B active, a more compact design. - GLM-4.5-Base and GLM-4.5-Air-Base: base models. - FP8 versions of the hybrid reasoning models are provided. - GLM-4.6: expands the context window from 128K to 200K tokens and reports improved coding, reasoning, tool-use and writing behavior. - GLM-4.7: adds Interleaved Thinking, Preserved Thinking and Turn-level Thinking, with reported gains on coding, terminal and tool-use benchmarks. - GLM-4.7-Flash: a lightweight 30B-A3B model for lighter deployment. Licensing and availability Models are released under the MIT license for commercial use and secondary development. Weights are distributed via Hugging Face and ModelScope. Model code, tool parsers and reasoning parsers are integrated into transformers, vLLM and SGLang. Deployment - Inference examples are given for transformers, vLLM and SGLang, including tool-call and reasoning parser flags (for example glm47 and glm45). - SGLang examples cover PD-disaggregation with prefill/decode separation and a router. - Thinking mode is enabled by default in vLLM and SGLang; it can be disabled per request via chat_template_kwargs. - Preserved Thinking for agentic tasks is configured through chat_template_kwargs (SGLang only). - Hardware tables list minimum and full-context GPU configurations for BF16 and FP8, for example GLM-4.5 FP8 on 8x H100 and GLM-4.5-Air FP8 on 2x H100, with larger counts for full 128K context. - Guides cover Ascend NPU deployment via xLLM and AMD GPU deployment. Fine-tuning Fine-tuning configurations are documented for Llama Factory and ms-swift, including LoRA, SFT and RL setups with H100 and H20 GPU counts. Evaluation and tooling The repository links a tool-integrated reasoning guide and a search-agent trajectory template, and provides an API request example for OpenAI-style tool calling. A technical report is available on arXiv, and API services are offered through the Z.ai platform.