GLM-5

2026-02-12 · Publisher : Z.ai · Domain : Artificial Intelligence

Z.ai has released GLM-5, a 744B-parameter MoE model (40B active) trained on 28.5T tokens, with DeepSeek Sparse Attention and asynchronous RL, under the MIT license, aimed at systems engineering and long-running agents.

### Context Z.ai has been listed on the Hong Kong Stock Exchange since January 8, 2026 (ticker 02513.HK, offer price 116.20 HKD per share according to the company's press release). GLM-4.7 dates from December 2025. In early 2026, the open models Kimi K2.5 and DeepSeek-V3.2 are the open references, and Claude Opus 4.5, Gemini 3 Pro and GPT-5.2 the closed references compared by the vendor. ### What the model brings - A jump from 355B (32B active) to 744B (40B active) and from 23T to 28.5T pre-training tokens. - The integration of DeepSeek Sparse Attention (DSA), which reduces deployment cost while keeping the long context. - slime, an asynchronous RL infrastructure that decouples generation and training to speed up post-training, with new agentic RL algorithms. - A goal of systems engineering and long agentic tasks, illustrated by Vending Bench 2 (Andon Labs): final balance of 4,432 USD, first among open models according to the vendor. - A clear improvement over GLM-4.7: SWE-bench Verified 77.8 versus 73.8; Terminal-Bench 2.0 (Terminus 2) 56.2 versus 41.0. ### Architecture and specifications - 744B in total, 40B active; about 753.9 billion parameters in the HF count (MTP layer included). DSA attention. - Context: 200K tokens; maximum output: 128K; modality: text; languages: Chinese and English; knowledge cutoff: not disclosed. - License: MIT. Weights: zai-org/GLM-5 and GLM-5-FP8. Report: "GLM-5: from Vibe Coding to Agentic Engineering" (arXiv 2602.15763, submitted February 17, 2026). - Deployment: vLLM, SGLang, KTransformers, Transformers, xLLM (Ascend); required hardware not quantified (startup example with tensor parallelism 8). ### Cost and access | Model (Z.ai API, current pricing) | Input /M | Cached input /M | Output /M | |---|---|---|---| | GLM-5 | 1 USD | 0.2 USD | 3.2 USD | ### Benchmarks | Benchmark | GLM-5 | Claude Opus 4.5 | Gemini 3 Pro | GPT-5.2 (xhigh) | |---|---|---|---|---| | SWE-bench Verified | 77.8 | 80.9 | 76.2 | 80.0 | | SWE-bench Multilingual | 73.3 | 77.5 | 65.0 | 72.0 | | Terminal-Bench 2.0 (Terminus 2) | 56.2 | 59.3 | 54.2 | 54.0 | | HLE with tools | 50.4 | 43.4* | 45.8* | 45.5* | | BrowseComp | 62.0 | 37.0 | 37.8 | - | | τ²-Bench | 89.7 | 91.6 | 90.7 | 85.5 | | GPQA-Diamond | 86.0 | 87.0 | 91.9 | 92.4 | | Vending Bench 2 (USD) | 4,432 | 4,967 | 5,478 | 3,591 | Source: GLM-5 Hugging Face model card (announcement table), scores measured by the vendor; asterisk: results on the full set. Vending Bench 2 is run by Andon Labs. ### Versus closed models In the vendor's table, GLM-5 is close to Claude Opus 4.5 in coding (77.8 versus 80.9 on SWE-bench Verified) and surpasses Gemini 3 Pro on SWE-bench Verified (76.2) and on SWE-bench Multilingual (65.0). It outperforms the three cited closed models on BrowseComp (62.0 versus 37.0, 37.8) and on HLE with tools (50.4 versus 43.4 to 45.8). In scientific reasoning, it is behind Gemini 3 Pro and GPT-5.2 on GPQA-Diamond (86.0 versus 91.9 and 92.4). On Vending Bench 2, the gap with Claude Opus 4.5 is 535 USD. The API is at 1 USD input and 3.2 USD output per million tokens; no closed-model pricing was verified to quantify the gap. The MIT license makes it possible to keep execution under control, but a 744B model requires a multi-GPU infrastructure in FP8. ### Key takeaways GLM-5 is the first GLM to exceed 700B parameters and to show near-parity with a top-tier closed model in agentic coding. It opens the GLM-5.x series, whose release cadence then accelerates.

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