# GLM-5.2 — The New Open-Source LLM That's Top of the AI Leaderboard

Z.ai releases GLM-5.2: a 753B parameter MoE model with MIT license, 1M-token context, and benchmark scores rivaling Claude Opus 4.8 and GPT-5.5. AIME 2026: 99.2.

## The New Champion

On June 16, 2026, **Z.ai** (formerly Zhipu AI) released **GLM-5.2** — an open-source large language model that has immediately reshuffled the AI leaderboard. With 753 billion total parameters operating through a Mixture-of-Experts architecture, GLM-5.2 achieves scores that match or exceed closed models like Claude Opus 4.8 and GPT-5.5 on key benchmarks.

What makes GLM-5.2 particularly significant is its **MIT license** — fully open, with no regional restrictions, allowing commercial use without any API key or usage limits. For teams that were forced to pay premium prices for Claude or GPT, GLM-5.2 offers a genuinely viable alternative.

## Benchmark Performance

### Reasoning

| **Benchmark** | **GLM-5.2** | **Opus 4.8** | **GPT-5.5** | **Gemini 3.1 Pro**|
--- | --- | --- | --- | ---
| **AIME 2026** | **99.2** | 95.7 | 98.3 | 98.2|
| HLE (w/ Tools) | 54.7 | 57.9* | 52.2* | 51.4*|
| **HMMT Feb 2026** | 92.5 | 96.7 | 96.7 | 87.3|
| GPQA-Diamond | 91.2 | 93.6 | 93.6 | 94.3|

GLM-5.2 **tops AIME 2026 at 99.2** — beating all closed models. On HLE with tools, Opus 4.8 leads slightly at 57.9, but GLM-5.2 holds strong at 54.7. The HMMT Feb 2026 math competition shows a wide gap between GLM-5.2 and Gemini (92.5 vs 87.3).

### Coding

| **Benchmark** | **GLM-5.2** | **Opus 4.8** | **GPT-5.5** | **Gemini 3.1 Pro**|
--- | --- | --- | --- | ---
| SWE-bench Pro | **62.1** | 69.2 | 58.6 | 54.2|
| Terminal Bench 2.1 | 81.0 | 85.0 | 84.0 | 74.0|
| **ProgramBench** | **63.7** | 71.9 | 70.8 | 39.5|
| **FrontierSWE** | **74.4** | 75.1 | 72.6 | 39.6|

GLM-5.2 **outperforms GPT-5.5 on SWE-bench Pro** (62.1 vs 58.6) and crushes Gemini 3.1 Pro on both ProgramBench (63.7 vs 39.5) and FrontierSWE (74.4 vs 39.6). While Opus 4.8 leads overall at 69.2, GLM-5.2's open license makes it more accessible for real-world deployment.

### Agentic / Long-Horizon

| **Benchmark** | **GLM-5.2** | **Opus 4.8** | **GPT-5.5** | **Gemini 3.1 Pro**|
--- | --- | --- | --- | ---
| SWE-Marathon | 13.0 | **26.0** | 12.0 | 4.0|
| PostTrainBench | 34.3 | **37.2** | 28.4 | 21.6|
| **MCP-Atlas** | **76.8** | 77.8 | 75.3 | 69.2|

Opus 4.8 leads on long-horizon tasks (SWE-Marathon, PostTrainBench), but GLM-5.2 **outperforms GPT-5.5 on MCP-Atlas** (76.8 vs 75.3). The long-horizon gap is real — both closed and open models struggle here.

### Key Takeaway

GLM-5.2 tops AIME 2026, matches Opus 4.8 on most benchmarks, beats GPT-5.5 on SWE-bench Pro, and crushes Gemini 3.1 Pro. With an MIT license and no usage restrictions, it's the highest-ranked open-source model on every benchmark listed.

## Technical Innovations

GLM-5.2 introduces several architectural innovations that enable its performance at 1M context:

- **IndexShare:** Reuses sparse attention indexer across every 4 transformer layers — reducing FLOPs by 2.9x at 1M context.

- **Enhanced MTP:** Multi-Token Prediction layer for speculative decoding — 20% longer acceptance length.

- **Effort Level Control:** Users can choose thinking depth (low/medium/high/max) balancing speed vs quality.

- **1M Context Training:** Trained from mid-training at 128K, fine-tuned on coding-agent trajectories.

## Deployment & Ecosystem

GLM-5.2 is designed for real-world deployment:

- Support frameworks: vLLM, SGLang, Transformers, Unsloth, KTransformers, Ascend NPU

- 282 safetensor files (large model, ~1.4TB storage)

- Tokenizer uses special tokens: `[gMASK]`, `<sop>`, `<|system|>`, `<|user|>`, `<|assistant|>`

- Available on HuggingFace: [zai-org/GLM-5.2](https://huggingface.co/zai-org/GLM-5.2) (19.7K+ downloads, 1.66K likes)

## The Chinese Open-Source Push

GLM-5.2's release represents a significant moment in the global AI landscape. Z.ai (formerly Zhipu AI, a Chinese company that went public on HKEX as 02513.HK) has produced the highest-ranked open-source model on every major benchmark, challenging the US dominance narrative in frontier AI.

Combined with recent events like the US export control suspension of Claude Fable 5 (June 12, 2026), open-source models like GLM-5.2 become even more critical — they offer unrestricted access to frontier-level intelligence without geopolitical dependencies.

## Related Articles

[**MiniMax M3**](/minimax-m3-frontier-coding-open-weight-model.html) — First open-weight frontier coding model with 1M context
[**Claude Fable 5 Suspended**](/claude-fable-5-suspended-us-government-export-control.html) — Why open-source matters more than ever