GLM-4.5: Agentic, Reasoning, and Coding (ARC) Foundation Models (Zhipu AI, 2025)

"GLM-4.5: Agentic, Reasoning, and Coding (ARC) Foundation Models" is the technical report for GLM-4.5 and its smaller sibling GLM-4.5-Air, published on arXiv (2508.06471) on August 8, 2025 by the GLM-4.5 Team of Zhipu AI and Tsinghua University. It describes an open-weight Mixture-of-Experts model with 355 billion total and 32 billion activated parameters, trained on 23 trillion tokens, that supports both a thinking mode for complex reasoning and agentic tasks and a non-thinking mode for immediate responses. The report organizes its claims around three capabilities it treats as the measure of a generalist model: agentic tool use, multi-step reasoning, and software-engineering coding ("ARC").

Architecture

GLM-4.5 is Zhipu's first MoE model. The report states that, unlike DeepSeek-V3 and Kimi K2, it reduces the model's width (hidden dimension and number of routed experts) and increases its depth, because "deeper models exhibited better reasoning capacity." It uses loss-free balance routing with sigmoid gates, Grouped-Query Attention with partial RoPE, QK-Norm to stabilize attention logits, and 2.5 times more attention heads than usual (96 heads for a 5,120 hidden dimension), which the authors report did not improve training loss but consistently improved reasoning benchmarks such as MMLU and BBH. An MoE layer serves as a Multi-Token Prediction layer to support speculative decoding.

Parameter counts include the MTP layers but exclude word embeddings and the output layer (Table 1).

Pre-training and mid-training

The pre-training corpus combines English and Chinese webpages, multilingual documents (from Zhipu's crawl and Fineweb-2), source code, and mathematics and science material. Webpages are bucketed by quality score, with the highest bucket contributing over 3.2 epochs and the lowest discarded; template-generated near-duplicates that MinHash deduplication missed are removed with SemDedup. Code is tiered by language-specific quality models, with low-quality code excluded and the Fill-In-the-Middle objective applied to all source code. A second pre-training stage up-samples code and coding, mathematics and science webpages.

"Mid-training" then adds medium-sized, domain-specific stages: repository-level code training with concatenated files and model-filtered GitHub issues, pull requests and commits (sequence length extended from 4K to 32K); synthetic reasoning data for mathematics, science and coding competitions; and long-context and agent training with sequence length extended to 128K and large-scale synthetic agent trajectories. The report's Figure 3 shows a 15T-token general pre-training corpus and a 7T-token code-and-reasoning corpus followed by mid-training stages of 500B, 500B and 100B tokens.

Training used the Muon optimizer for all parameters except word embeddings, biases and RMSNorm weights, which the authors report accelerated convergence and tolerated larger batch sizes, with a cosine learning-rate schedule (warm-up to 2.5e-4, decay to 2.5e-5) after early experiments found a warmup-stable-decay schedule underfit on general benchmarks. Batch size was warmed up from 16M to 64M tokens over the first 500B tokens.

Post-training

Post-training runs in two stages: expert training, which builds separate Reasoning, Agent and General-chat expert models, and unified training, which uses self-distillation to merge them into one hybrid model that can either reason deliberately or answer directly. Supervised fine-tuning provides a cold start for each expert and later distills the experts into the unified model, with training data balanced between responses with and without explicit reasoning. The report describes:

• A function-call template that places argument keys and values in XML-like special tokens rather than JSON, so that code passed as an argument does not need extensive character escaping.
• Data selection. Removing the easiest 50% of prompts by response length improved mathematics and science results by 2–4% while halving the data; generating four responses per hard prompt added a further 1–2%.
• Automatic agentic SFT data built from collected agent frameworks, real tool APIs and MCP servers, and LLM-simulated tools, with synthesized tasks, generated trajectories and multiple judge agents keeping only successful trajectories.
• Reasoning RL built on GRPO without the KL loss term, with a two-stage difficulty-based curriculum (switching to problems unsolved at pass@8 but solved at pass@512), and a single RL stage at the full 64K output length, which the authors found outperformed progressively lengthened stages because shorter stages caused the model to "unlearn" long outputs. These ablations were run on a smaller experimental model.
• Agentic RL on web search and software engineering with iterative self-distillation, and the observation that accuracy on BrowseComp scales smoothly with the number of interaction turns, a form of test-time compute that scales interaction rather than output tokens.
• General RL combining rule-based feedback, human feedback (RLHF) and model-based feedback over roughly 5,000 balanced prompts, instruction-following RL against a taxonomy of 7 major and 151 minor constraint types (the authors report no clear evidence of reward hacking up to about 1,000 steps), and function-calling RL with a strict all-or-nothing reward for each step plus end-to-end multi-turn RL.

The RL infrastructure is Slime, an open-source framework the team developed (Megatron for training, SGLang for rollouts, a data buffer between them), which supports a synchronous colocated mode for reasoning tasks and a disaggregated asynchronous mode for long-running agentic environments such as software engineering.

Availability

Both models were released with open weights on Hugging Face and are served through Z.ai and BigModel.cn; the team also released an evaluation toolkit (glm-simple-evals) for reproducing the benchmark results, and the CC-Bench task descriptions and trajectories.

Reported results

Evaluation results the document states for which there is no separate benchmark page; results on benchmarks that have a page are recorded there and cite this page. Scores and settings are as printed in the document.

Provenance

Full text extracted from the arXiv PDF (2508.06471) via bin/fetch-source.py and held at Raw Sources/GLM-4.5 (2025).md. The extraction de-spaced some passages and flattened several results tables (Tables 3–6 and 8–11); figures above are taken from the report's prose and from tables whose extracted values are unambiguous. All benchmark results are the developer's own and are not independently reproduced here.

Relationships

• supports: GLM-4 generation (GLM-4 through GLM-4.7) — the technical documentation for the release
• related: GLM-5: from Vibe Coding to Agentic Engineering (Zhipu AI, 2026) — the successor report
• instance-of: Open-Weight Frontier Models — an open-weight model reported at frontier-comparable scores
• related: Mixture of Experts (MoE), Reasoning Models and Chain-of-Thought (hybrid thinking and non-thinking modes), RLHF (Reinforcement Learning from Human Feedback), Reward Hacking
• related: Zhipu AI, US-China AI Competition: Different Races, Different Metrics