Zhipu: glm-5-thinking

We are launching GLM-5, targeting complex systems engineering and long-horizon agentic tasks. Scaling is still one of the most important ways to improve the intelligence efficiency of Artificial General Intelligence (AGI). Compared to GLM-4.5, GLM-5 scales from 355B parameters (32B active) to 744B parameters (40B active), and increases pre-training data from 23T to 28.5T tokens. GLM-5 also integrates DeepSeek Sparse Attention (DSA), significantly reducing deployment cost while preserving long-context capacity. Reinforcement learning aims to bridge the gap between competence and excellence in pre-trained models. However, deploying it at scale for LLMs is a challenge due to RL training inefficiency. To this end, we developed slime, a novel asynchronous RL infrastructure that substantially improves training throughput and efficiency, enabling more fine-grained post-training iterations. With advances in both pre-training and post-training, GLM-5 delivers significant improvement compared to GLM-4.7 across a wide range of academic benchmarks and achieves best-in-class performance among all open-source models in the world on reasoning, coding, and agentic tasks, closing the gap with frontier models. Image 1 GLM-5 is designed for complex systems engineering and long-horizon agentic tasks. On our internal evaluation suite CC-Bench-V2, GLM-5 significantly outperforms GLM-4.7 across frontend, backend, and long-horizon tasks, narrowing the gap to Claude Opus 4.5. Image 2 On Vending Bench 2, a benchmark that measures long-term operational capability, GLM-5 ranks #1 among open-source models. Vending Bench 2 requires the model to run a simulated vending machine business over a one-year horizon; GLM-5 finishes with a final account balance of $4,432, approaching Claude Opus 4.5 and demonstrating strong long-term planning and resource management. Image 3 GLM-5 is open-sourced on Hugging Face and ModelScope, with model weights released under the MIT License. GLM-5 is also available on developer platform api.z.ai and BigModel.cn, with compatibility with Claude Code and OpenClaw. You can also try it for free on Z.ai. | Benchmark | GLM-5 (Thinking) | GLM-4.7 (Thinking) | DeepSeek-V3.2 (Thinking) | Kimi K2.5 (Thinking) | Claude Opus 4.5 (Extend Thinking) | Gemini 3.0 Pro (High Thinking Level) | GPT-5.2 (xhigh) | | --- | --- | --- | --- | --- | --- | --- | --- | | Reasoning | | Humanity's Last Exam | 30.5 | 24.8 | 25.1 | 31.5 | 28.4 | 37.2 | 35.4 | | Humanity's Last Exam w/ Tools | 50.4 | 42.8 | 40.8 | 51.8 | 43.4 | 45.8 | 45.5* | | AIME 2026 I | 92.7 | 92.9 | 92.7 | 92.5 | 93.3 | 90.6 | - | | HMMT Nov. 2025 | 96.9 | 93.5 | 90.2 | 91.1 | 91.7 | 93.0 | 97.1 | | IMOAnswerBench | 82.5 | 82.0 | 78.3 | 81.8 | 78.5 | 83.3 | 86.3 | | GPQA-Diamond | 86.0 | 85.7 | 82.4 | 87.6 | 87.0 | 91.9 | 92.4 | | Coding | | SWE-bench Verified | 77.8 | 73.8 | 73.1 | 76.8 | 80.9 | 76.2 | 80.0 | | SWE-bench Multilingual | 73.3 | 66.7 | 70.2 | 73.0 | 77.5 | 65.0 | 72.0 | | Terminal-Bench 2.0 Terminus-2 | 56.2 / 60.7† | 41.0 | 39.3 | 50.8 | 59.3 | 54.2 | 54.0 | | Terminal-Bench 2.0 Claude Code | 56.2 / 61.1† | 32.8 | 46.4 | - | 57.9 | - | - | | CyberGym | 43.2 | 23.5 | 17.3 | 41.3 | 50.6 | 39.9 | - | | General Agent | | BrowseComp | 62.0 | 52.0 | 51.4 | 60.6 | 37.0 | 37.8 | - | | BrowseComp w/ Context Manage | 75.9 | 67.5 | 67.6 | 74.9 | 67.8 | 59.2 | 65.8 | | BrowseComp-Zh | 72.7 | 66.6 | 65.0 | 62.3 | 62.4 | 66.8 | 76.1 | | τ²-Bench | 89.7 | 87.4 | 85.3 | 80.2 | 91.6 | 90.7 | 85.5 | | MCP-Atlas Public Set | 67.8 | 52.0 | 62.2 | 63.8 | 65.2 | 66.6 | 68.0 | | Tool-Decathlon | 38.0 | 23.8 | 35.2 | 27.8 | 43.5 | 36.4 | 46.3 | | Vending Bench 2 | $4,432.12 | $2,376.82 | $1,034.00 | $1,198.46 | $4,967.06 | $5,478.16 | $3,591.33 | *: refers to their scores of full set. †: A verified version of Terminal-Bench 2.0 that fixes some ambiguous instructions. See footnote for more evaluation details. Office Foundation models are moving from “chat” to “work,” much like Office tools for knowledge workers and programming tools for engineers. GLM-4.5 is our first step for reasoning, coding, and agent, enabling the model to complete complex tasks. With GLM-5, we further enhance complex systems engineering and long-horizon agent capabilities. GLM-5 can turn text or source materials directly into .docx, .pdf, and .xlsx files—PRDs, lesson plans, exams, spreadsheets, financial reports, run sheets, menus, and more—delivered end-to-end as ready-to-use documents. Our official application, Z.ai is rolling out an Agent mode with built-in skills for PDF / Word / Excel creation, supporting multi-turn collaboration and turning outputs into real deliverables. Document (.docx) generated by GLM-5 Image 4 Getting started with GLM-5 Use GLM-5 with GLM Coding Plan Try GLM-5 in your favorite coding agents—Claude Code, OpenCode, Kilo Code, Roo Code, Cline, Droid, and more. https://docs.z.ai/devpack/overview For GLM Coding Plan subscribers: Due to limited compute capacity, we’re rolling out GLM-5 to Coding Plan users gradually. Max plan users: You can enable GLM-5 now by updating the model name to "GLM-5" (e.g. in ~/.claude/settings.json for Claude Code). Other plan tiers: Support will be added progressively as the rollout expands. Quota note: Requests to GLM-5 consume more plan quota than GLM-4.7. Prefer a GUI? We offer Z Code —an agentic development environment that lets you control (even remotely) multiple agents and have them collaborate on complex tasks. Start building now:https://z.ai/subscribe Use GLM-5 with OpenClaw Beyond coding agents, GLM-5 also supports OpenClaw—a framework that turns GLM-5 into a personal assistant that can operate across apps and devices, not just chat. OpenClaw is included in GLM Coding Plan. See the guidance. Chat with GLM-5 on Z.ai GLM-5 is accessible through Z.ai. Manually change the model option to GLM-5, if the system does not automatically do so. We offer both Chat and Agent mode for GLM-5: Chat Mode: Instant response, interactive chat, lightweight delivery Agent Mode: Multiple tools, diverse skills, delivering results directly Serve GLM-5 Locally The model weights of GLM-5 are publicly available on HuggingFace and ModelScope. For local deployment, GLM-5 supports inference frameworks including vLLM and SGLang. Comprehensive deployment instructions are available at the official GitHub repository. We also support deploying GLM-5 on non-NVIDIA chips, including Huawei Ascend, Moore Threads, Cambricon, Kunlun Chip, MetaX, Enflame, and Hygon. Through kernel optimization and model quantization, GLM-5 can achieve a reasonable throughput on those chips. Footnote Humanity’s Last Exam (HLE) & other reasoning tasks: We evaluate with a maximum generation length of 131,072 tokens (temperature=1.0, top_p=0.95, max_new_tokens=131072). By default, we report the text-only subset; results marked with are from the full set. We use GPT-5.2 (medium) as the judge model. For HLE-with-tools, we use a maximum context length of 202,752 tokens. SWE-bench & SWE-bench Multilingual: We run the SWE-bench suite with OpenHands using a tailored instruction prompt. Settings: temperature=0.7, top_p=0.95, max_new_tokens=16384, with a 200K context window. BrowserComp: Without context management, we retain details from the most recent 5 turns. With context management, we use the same discard-all strategy as DeepSeek-V3.2 and Kimi K2.5. Terminal-Bench 2.0 (Terminus 2): We evaluate with the Terminus framework using timeout=2h, temperature=0.7, top_p=1.0, max_new_tokens=8192, with a 128K context window. Resource limits are capped at 16 CPUs and 32 GB RAM. CyberGym: We evaluate in Claude Code 2.1.18 (think mode, no web tools) with (temperature=1.0, top_p=1.0, max_new_tokens=32000) and a 250-minute timeout per task. Results are single-run Pass@1 over 1,507 tasks. MCP-Atlas: All models are evaluated in think mode on the 500-task public subset with a 10-minute timeout per task. We use Gemini 3 Pro as the judge model. τ²-bench: We add a small prompt adjustment in Retail and Telecom to avoid failures caused by premature user termination. For Airline, we apply the domain fixes proposed in the Claude Opus 4.5 system card. Vending Bench 2: Runs are conducted independently by Andon Labs.