GLM-5.2
Open Weights
Zhipu AI
Zhipu's 753B-parameter flagship, built for long-horizon agentic work and released under the MIT licence with no regional restrictions. It sustains a 1M-token context - a five-fold jump over GLM-5.1's 200K - using an IndexShare architecture that reuses one indexer across every four sparse-attention layers, cutting per-token FLOPs by 2.9x at full context length. A reasoning_effort parameter selects between 'high' and 'max' thinking budgets.
Strengths
- MIT licensed with no regional limits - among the most permissive frontier-scale releases
- 1M-token context, five times GLM-5.1's 200K window
- IndexShare architecture cuts per-token FLOPs by 2.9x at 1M context
- 99.2 on AIME 2026 and 91.2 on GPQA-Diamond
- 62.1 on SWE-Bench Pro and 81.0 on Terminal Bench 2.1
- reasoning_effort parameter trades latency against depth ('high' or 'max')
Caveats
- Text only - no vision or audio input
- 753B parameters means self-hosting needs substantial GPU capacity
- No release date is stated on the model card
- Tool-Decathlon score of 48.2 is well below its reasoning results, so agentic breadth lags raw reasoning
- Vendor
- Zhipu AI
- Context Window
- 1,000,000 tokens
- License
- MIT
- Modalities
- text
| Benchmark | Score | Max |
|---|---|---|
| AIME 2026 | 99.2 | 100.0 |
| GPQA-Diamond | 91.2 | 100.0 |
| Terminal Bench 2.1 | 81.0 | 100.0 |
| MCP-Atlas | 76.8 | 100.0 |
| SWE-Bench Pro | 62.1 | 100.0 |
| Tool-Decathlon | 48.2 | 100.0 |
| HLE | 40.5 | 100.0 |
Capabilities
Vision
Audio
Video
Tool Use
Resources
Pricing
Open weights under the MIT licence with no regional limits - self-host at no licence cost. Also served through Z.ai's API.