GLM-4.5 and GLM-4.5-Air

2025-07-28 · Publisher : Z.ai · Domain : Artificial Intelligence

Z.ai (formerly Zhipu AI) has released GLM-4.5 (355B parameters, 32B active) and GLM-4.5-Air (106B, 12B active), two MoE models with hybrid reasoning for agents and code, under the MIT license.

### Context In mid-2025, Chinese open MoE models (DeepSeek-V3/R1, Kimi K2, Qwen3) compete for the top of the open ranking, while closed models still dominate agentic coding (Claude 4, o3, Gemini 2.5). Zhipu AI has just released GLM-4-0414 and GLM-4.1V-Thinking. GLM-4.5 is the vendor's first model to explicitly target reasoning, code and agents all together in a single model. ### What the model brings - Hybrid reasoning: a thinking mode for complex reasoning and tools, a non-thinking mode for immediate answers. - A MoE architecture that is deeper than it is wide, with MTP layers for speculative decoding. - Multi-stage training on 23T tokens (15T of general corpus then 7T of code and reasoning, according to the post), followed by expert iterations and reinforcement learning. - Native agentic capabilities: function calling, web browsing (BrowseComp), full-stack development, presentation generation. - Announced compatibility with the Claude Code framework and CC-Bench evaluations on 52 coding tasks. ### Architecture and specifications - GLM-4.5: 355B total, 32B active. GLM-4.5-Air: 106B total, 12B active. Base, hybrid and FP8 versions published. - Context: 128K tokens; maximum output: 96K (API documentation). Modality: text. Languages: Chinese and English. - Knowledge cutoff: not disclosed. - License: MIT, commercial use and derivatives allowed. - Weights: Hugging Face (zai-org/GLM-4.5, GLM-4.5-Air) and ModelScope. Technical report: arXiv 2508.06471. ### Cost and access | Model (Z.ai API, current pricing) | Input /M | Output /M | Cached input /M | |---|---|---|---| | GLM-4.5 | 0.6 USD | 2.2 USD | 0.11 USD | | GLM-4.5-Air | 0.2 USD | 1.1 USD | 0.03 USD | | GLM-4.5-X | 2.2 USD | 8.9 USD | 0.45 USD | Self-hosting (model card, with MTP, batch size less than or equal to 8): GLM-4.5 in FP8 on 8 H100 or 4 H200; GLM-4.5-Air in FP8 on 2 H100 or 1 H200; full 128K context: 16 H100 (GLM-4.5 FP8) or 4 H100 (Air FP8). ### Benchmarks | Measure | GLM-4.5 | Comparison | |---|---|---| | BrowseComp (with browsing) | 26.4% | Claude-4-Opus 18.8%; o4-mini-high 28.3% | | SWE-bench Verified | 64.2% | Closed model scores not included (table in image) | | TAU-Bench | 70.1% | same | | AIME 24 | 91.0% | same | | Average over 12 benchmarks | 63.2 (3rd rank) | GLM-4.5-Air: 59.8 | Sources: official post z.ai/blog/glm-4.5 (BrowseComp), summary of the arXiv 2508.06471 report (SWE-bench, TAU, AIME) and Hugging Face model card (average); scores measured by the vendor. On CC-Bench, GLM-4.5 wins 53.9% of the time against Kimi K2 and 80.8% against Qwen3-Coder, with "room for improvement" against Claude-4-Sonnet according to the vendor (no figure given). ### Versus closed models According to the vendor, GLM-4.5 ranks 3rd out of 12 benchmarks among all the models evaluated, open and closed, and 2nd on agentic benchmarks (arXiv report). On BrowseComp, it outperforms Claude-4-Opus (26.4% versus 18.8%) and remains close to o4-mini-high (28.3%). The vendor acknowledges a lag behind Claude-4-Sonnet in agentic coding (CC-Bench), without giving the score. The API is at 0.6 USD for input and 2.2 USD for output per million tokens, a level that the vendor presents as very low; no contemporary closed model pricing was verified to quantify the gap. For enterprises, the MIT license and the reduced hardware footprint (8 H100 in FP8) make self-hosting a frontier-class model conceivable. ### Key takeaways GLM-4.5 is the milestone that takes the GLM models from 32B models to a 355B MoE, under MIT and with a complete agentic mode. The Air version (106B, 12B active) is the most accessible for self-hosting.

Tags : LLM, Open Source, Z.ai, Zhipu AI, GLM, MIT, MoE, Agents