GLM-4.7 (and GLM-4.7-Flash)

2025-12-22 · Publisher : Z.ai · Domain : Artificial Intelligence

Z.ai has released GLM-4.7, an open coding-oriented model (73.8% on SWE-bench Verified), with thinking between actions, thinking preserved across turns and per-turn control of reasoning, under the MIT license.

Z.ai has released GLM-4.7, an open coding-oriented model (73.8% on SWE-bench Verified), with thinking between actions, thinking preserved across turns and per-turn control of reasoning, under the MIT license. ### Context In late 2025, the leading closed models are Gemini 3 Pro, Claude Opus 4.5 and GPT-5.1, and agentic coding remains their strong point. GLM-4.7 follows GLM-4.6 three months later, with an explicit goal of being a "coding partner". On January 19, 2026, GLM-4.7-Flash, a 30B-A3B MoE, rounds out the lineup for lightweight deployment. ### What the model brings - +5.8 points on SWE-bench Verified (73.8 versus 68.0) and +12.9 points on SWE-bench Multilingual (66.7 versus 53.8) compared with GLM-4.6. - +16.5 points on Terminal Bench 2.0 (41.0 versus 24.5) and 42.8% on HLE with tools (+12.4 points). - Interleaved Thinking, plus two new features: Preserved Thinking (thinking blocks are kept from one turn to the next in coding agents) and Turn-level Thinking (thinking can be enabled per turn to trade off latency and accuracy). - Improved interface quality: cleaner web pages, better-sized slides. ### Architecture and specifications - About 358.3 billion parameters by the HF count (GLM-4.5 family, MoE); number of active parameters not included in the model card. - Context: 200K tokens; maximum output: 128K (API documentation). Modality: text. Knowledge cutoff: not disclosed. - GLM-4.7-Flash: 30B-A3B MoE (31.2 billion by the HF count), 200K context. - License: MIT. Weights: zai-org/GLM-4.7, GLM-4.7-FP8, GLM-4.7-Flash. No dedicated report: refers to arXiv 2508.06471. ### Cost and access | Model (Z.ai API, current pricing) | Input /M | Output /M | Cache /M | |---|---|---|---| | GLM-4.7 | 0.6 USD | 2.2 USD | 0.11 USD | | GLM-4.7-FlashX | 0.07 USD | 0.4 USD | 0.01 USD | | GLM-4.7-Flash | Free | Free | Free | Self-hosting: vLLM and SGLang (main branches at release); required hardware not disclosed. ### Benchmarks | Benchmark | GLM-4.7 | Claude Sonnet 4.5 | GPT-5-High | GPT-5.1-High | Gemini 3.0 Pro | |---|---|---|---|---|---| | SWE-bench Verified | 73.8 | 77.2 | 74.9 | 76.3 | 76.2 | | Terminal Bench 2.0 | 41.0 | 42.8 | 35.2 | 47.6 | 54.2 | | GPQA-Diamond | 85.7 | 83.4 | 85.7 | 88.1 | 91.9 | | AIME 2025 | 95.7 | 87.0 | 94.6 | 94.0 | 95.0 | | HLE with tools | 42.8 | 32.0 | 35.2 | 42.7 | 45.8 | | τ²-Bench | 87.4 | 87.2 | 82.4 | 82.7 | 90.7 | | BrowseComp | 52.0 | 24.1 | 54.9 | 50.8 | - | Source: Hugging Face model card for GLM-4.7, scores published and measured by the vendor. ### Versus closed models In the vendor's table, GLM-4.7 is slightly below GPT-5-High on SWE-bench Verified (73.8 versus 74.9) but ahead of it on HLE with tools (42.8 versus 35.2), τ²-Bench (87.4 versus 82.4) and Terminal Bench 2.0 (41.0 versus 35.2). It stays behind Gemini 3.0 Pro on 5 of the 6 comparable rows (for example 76.2 versus 73.8 on SWE-bench Verified and 54.2 versus 41.0 on Terminal Bench 2.0), except AIME 2025 (95.7 versus 95.0), and behind GPT-5.1-High on Terminal Bench 2.0 (47.6). The comparison also includes Claude Sonnet 4.5, ahead in coding (77.2 on SWE-bench Verified). The API price (0.6 USD input, 2.2 USD output) and the MIT license are the differentiating points; the agentic coding score gap with the top of the closed ranking remains 2 to 13 points on these rows. ### Key takeaways GLM-4.7 narrows the coding gap with the closed models of late 2025 without closing it. Preserved Thinking, useful in long coding agents, is the main new feature in practice.

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