On April 7, Z.ai (formerly Zhipu AI) released GLM-5.1, a 754B-parameter MoE architecture model open-sourced under the MIT license.
Benchmark results: SWE-Bench Pro 58.4%, first globally. GPT-5.4 scored 57.7%, and Claude Opus 4.6 scored 57.3%.
That number alone is enough to generate buzz — an open-source model has directly surpassed two closed-source flagships in coding capability. But there's something even more interesting:
The model was trained entirely on Huawei Ascend 910B chips and the MindSpore framework, with no NVIDIA GPUs involved.
Technical Specifications
First, the hard specs:
| Parameter | Value |
| Total Parameters | 754B |
| Active Parameters per Inference | 40B |
| Architecture | MoE + Dynamic Sparse Attention |
| Context Window | 200K tokens |
| Maximum Output | 131,072 tokens |
| Model File Size | 1.51TB (HuggingFace) |
| License | MIT |
Key benchmark scores:
- SWE-Bench Pro: 58.4% (first globally)
- CyberGym: 68.7%
- BrowseComp: 68.0%
- AIME 2026: 95.3%
SWE-Bench Pro tests real GitHub issue fixes, not synthetic datasets that can be gamed. The 58.4% figure carries real weight.
An Agent That Runs for 8 Hours
GLM-5.1 has a less-discussed but important capability: autonomous execution.
Z.ai describes it as "independently executing tasks for over 8 hours, spanning hundreds of rounds of operations and thousands of tool calls" — it can work continuously for more than 8 hours, handling hundreds of task rounds and thousands of tool calls.
Most models break down after a few dozen rounds in complex agent tasks: the context window fills up, reasoning degrades, and tool calls start looping. GLM-5.1 clearly has specialized optimizations here, not just relying on a large context window.
This is a substantive advance for enterprise-level agent deployment, not just a benchmark score.
The Key Point: No NVIDIA
Training hardware: Huawei Ascend 910B
Training framework: Huawei MindSpore
This is not a minor detail.
For a long time, the optimal path for training large-scale models has been almost exclusively NVIDIA CUDA. H100, H800, A100 — use whichever is available. MindSpore has consistently lagged behind PyTorch in engineering maturity.
GLM-5.1 proves this path is viable: not just that it runs, but that it produced a model that beats GPT-5.4 and Claude Opus 4.6.
"GLM-5.1 Tops SWE-Bench Pro With Zero NVIDIA Hardware"
— Awesome Agents
The political significance of this is as large as the technical significance. US export controls on AI chips to China are now a real production constraint, not a future threat. GLM-5.1 is an experimental result: what Chinese AI teams can achieve under compute constraints.
The answer so far: first globally on SWE-Bench Pro, open-sourced under MIT, API pricing at $1.26/M tokens.
Price Comparison
GLM-5.1 API pricing:
- Input: $1.26/M tokens
- Output: $3.96/M tokens
For comparison: Claude Opus 4.6 is roughly $15/$75, and GPT-5.4 is in the same range.
With SWE-Bench Pro's top coding capability at this price point, the options are clear for AI coding infrastructure developers.
Of course, SWE-Bench Pro doesn't represent all scenarios. GLM-5.1 still lags behind top closed-source models in mathematical reasoning and general benchmarks, and the original article doesn't avoid this. It's not an all-around champion — it's a specialized model differentiated in coding and agent directions.
A Trend Worth Watching
In the past year, Chinese open-source models including Z.ai (GLM-5.1), DeepSeek (V3.2), and Kimi K2 have topped various specialized benchmarks, using MIT or Apache licenses, with API prices crushing closed-source competitors.
Meta has announced partial closure, and OpenAI has never truly open-sourced its flagships — yet Chinese companies continue to release weights publicly.
This landscape was almost unimaginable in early 2025.
Sources: Z.ai Releases GLM-5.1: 754B Model Tops SWE-Bench Pro (WinBuzzer); CocoLoop, GLM-5.1 Tops SWE-Bench Pro With Zero NVIDIA Hardware (Awesome Agents); GLM-5.1 vs Claude, GPT, Gemini, DeepSeek: how Zhipu AI's model stacks up (APIdog)