GLM-5.1 is an open-source flagship AI model developed by Z.ai (formerly Zhipu AI, a Tsinghua University spinoff and the first publicly traded foundation model company). Released on April 7, 2026, it is a post-training upgrade to GLM-5, built on a 754-billion parameter Mixture-of-Experts architecture with 40 billion active parameters per token. GLM-5.1 is purpose-built for agentic engineering and long-horizon autonomous software development, scoring 58.4% on SWE-Bench Pro (ranked #1 as of April 2026).
- Agentic Coding: SWE-Bench Pro 58.4% (ranked #1 as of April 2026), ahead of GPT-5.4 (57.7%) and Claude Opus 4.6 (57.3%).
- 8-Hour Autonomous Execution: Can work autonomously on a single task for up to 8 hours, completing full plan-execute-test-fix-optimize loops across hundreds of iterations and thousands of tool calls.
- MIT Licensed Open Weights: Released under the MIT license on Hugging Face — one of the most permissive open-source licenses, allowing unrestricted commercial use, modification, and fine-tuning.
- Ascend-Native Training: Trained entirely on Huawei Ascend 910B chips using the MindSpore framework, achieving full independence from US-manufactured hardware.
- Long-Horizon Software Engineering: Excels at complex, multi-step coding tasks requiring sustained autonomous execution — e.g., building a complete Linux desktop system from scratch within 8 hours.
- Agentic Tool Orchestration: Strong function calling, MCP integration, and structured output support make it ideal for building AI agents that need to interact with external tools and APIs.
- Cost-Effective Frontier Performance: At $1.40/$4.40 per million input/output tokens, it delivers ~94.6% of Claude Opus 4.6's coding capability at roughly 5-8x lower cost.
| Capability | Description |
|---|
| Reasoning | AIME 2026: 95.3%, GPQA-Diamond: 86.2%, strong system-level reasoning across planning and iterative debugging |
| Coding | SWE-Bench Pro 58.4% (SOTA), CyberGym 68.7%, BrowseComp 68.0%, MCP-Atlas 71.8% |
| Multimodal | Text only. No image, audio, or video input (separate GLM-5V-Turbo variant available for vision) |
| Response Speed | Not independently benchmarked yet; comparable to similar-scale MoE models |
| Context Window | 200K tokens |
| Max Output | 128K tokens |
| Tool Use | Function calling, structured output, context caching, MCP integration, thinking mode |
| Multilingual | Strong multilingual support, particularly Chinese and English |
- Text-only input; no native multimodal support (vision is handled by the separate GLM-5V-Turbo model).
- Math and science scores (AIME 2026: 95.3%, GPQA-Diamond: 86.2%) trail top proprietary models, making it a weaker choice for pure quantitative research.
- On broader coding composites (Terminal-Bench 2.0 + NL2Repo), Claude Opus 4.6 still leads at 57.5 vs GLM-5.1's 54.9.
- Self-hosting requires significant compute resources due to the 754B parameter count.
| Model | Input (Credits/Token) | Cache Write (Credits/Token) | Cache Read (Credits/Token) | Output (Credits/Token) |
|---|
| GLM-5.1 | 1.40 | 1.40 | 0.26 | 4.40 |
