GLM-5.2 is the flagship open-source GLM AI model: with a stable context of one million tokens, non-thinking/medium/high/three levels of thinking intensity, it reports scores of 81.0 on Terminal-Bench 2.1 and 74.4 on FrontierSWE. Below, you can compare each GLM model — or experience the free AI sandbox first.
No registration required, no account required, no quota required. Real-time streaming responses can be obtained by entering prompt words.
The reply will be streamed out here...
This experience uses the GLM large language model, and real-time streaming responses can be obtained by entering prompt words.
Four capabilities define the current generation of GLM models — all reported by the model authors and reproducible from open rights.
A context of one million tokens is not only available but also maintains quality: GLM-5.2 has been trained on long context encoding agent trajectories such as large-scale implementation, automated research, and complex debugging.
Non-thinking, high, and highest three levels allow you to weigh between delay and capability for tasks — on the agent programming evaluation, about 63% (~35K output tokens) to 74% (~83K).
Every four sparse attention layers share a lightweight indexer, reducing the FLOPs per token by 2.9 times under a context length of one million, while maintaining long-distance quality without loss.
Open-source GLM models under the MIT license: weights are published on HuggingFace and ModelScope, without regional restrictions, and support local deployment through transformers, vLLM, SGLang, xLLM, and ktransformers.
All the following data are the results of GLM-5.2 and comparative models published by the model authors. glmmodel.net only provides reports and does not run these evaluations.
Task duration can reach 20 hours
Up to 10 hours on a single H100 card
Ultra-long cycle software engineering, up to 10 hours
In three long-term benchmark tests, GLM-5.2 is the highest-ranking open-source model — proving that its million-token context can be converted into actual output, not just an acceptable number of tokens. On FrontierSWE, it is only one point behind Opus 4.8 and one point ahead of GPT-5.5, and eleven points ahead of the previous generation Opus 4.7.
Published GLM model benchmark results, covering 17 evaluations of 8 models.
| Benchmark Test | GLM-5.2 | GLM-5.1 | Qwen3.7-Max | MiniMax M3 | DeepSeek-V4-Pro | Claude Opus 4.8 | GPT-5.5 | Gemini 3.1 Pro |
|---|---|---|---|---|---|---|---|---|
| Inference | ||||||||
| HLE | 40.5 | 31.0 | 41.4 | 37.0 | 37.7 | 49.8* | 41.4* | 45.0 |
| HLE (With Tools) | 54.7 | 52.3 | 53.5 | – | 48.2 | 57.9* | 52.2* | 51.4* |
| CritPt | 20.9 | 4.6 | 13.4 | 3.7 | 12.9 | 20.9 | 27.1 | 17.7 |
| AIME 2026 | 99.2 | 95.3 | 97.0 | – | 94.6 | 95.7 | 98.3 | 98.2 |
| HMMT November 2025 | 94.4 | 94.0 | 95.0 | 84.4 | 94.4 | 96.5 | 96.5 | 94.8 |
| HMMT February 2026 | 92.5 | 82.6 | 97.1 | 84.4 | 95.2 | 96.7 | 96.7 | 87.3 |
| IMOAnswerBench | 91.0 | 83.8 | 90.0 | – | 89.8 | 83.5 | – | 81.0 |
| GPQA-Diamond | 91.2 | 86.2 | 90.0 | 93.0 | 90.1 | 93.6 | 93.6 | 94.3 |
| Programming | ||||||||
| SWE-bench Pro | 62.1 | 58.4 | 60.6 | 59.0 | 55.4 | 69.2 | 58.6 | 54.2 |
| NL2Repo | 48.9 | 42.7 | 47.2 | 42.1 | 35.5 | 69.7 | 50.7 | 33.4 |
| DeepSWE | 46.2 | 18.0 | 18.0 | 20.0 | 8.0 | 58.0 | 70.0 | 10.0 |
| ProgramBench | 63.7 | 50.9 | – | – | 47.8 | 71.9 | 70.8 | 39.5 |
| Terminal-Bench 2.1 (Terminus-2) | 81.0 | 63.5 | 75.0 | 65.0 | 64.0 | 85.0 | 84.0 | 74.0 |
| Terminal-Bench 2.1 (Best Framework) | 82.7 (Claude Code) | 69.0 (Claude Code) | – | – | – | 78.9 (Claude Code) | 83.4 (Codex) | 70.7 (Gemini CLI) |
| FrontierSWE Dominance | 74.4 | 30.5 | – | – | 29.0 | 75.1 | 72.6 | 39.6 |
| PostTrainBench | 34.3 | 20.1 | – | – | – | 37.2 | 28.4 | 21.6 |
| SWE-Marathon | 13.0 | 1.0 | – | – | – | 26.0 | 12.0 | 4.0 |
| Agent | ||||||||
| MCP-Atlas (public dataset) | 76.8 | 71.8 | 76.4 | 74.2 | 73.6 | 77.8 | 75.3 | 69.2 |
| Tool-Decathlon | 48.2 | 40.7 | – | – | 52.8 | 59.9 | 55.6 | 48.8 |
* Scores are based on complete datasets.
GLM models are not just one speed. Think force control allows you to allocate more computing resources only when needed.
~35K average tokens
Scenarios where latency is more important than depth, such as quick editing, template code, and routine refactoring.
~43K average tokens
The default choice for most agent programming: approaching the highest quality with about half the token consumption.
~83K average tokens
Difficult long-term problems — allocate additional computing resources only when the task truly requires it.
The data comes from the average of Terminal-Bench 2.1, DeepSWE, and SWE-Atlas QnA, evaluated on Claude Code 2.1.167. Reported by the model authors. Under similar token budgets, GLM-5.2's performance is between Claude Opus 4.7 and Opus 4.8.
Increasing the context limit from 200K to one million tokens is an engineering issue rather than a configuration parameter. Three modifications are key.
Every four Transformer layers share a lightweight indexer. It is located at the first layer among the four, and its top-k index is reused by the other three layers — eliminating the dot product and top-k calculation in three-quarters of the layers. Result: Each token's FLOPs is reduced by 2.9 times under a context length of one million. IndexShare was introduced starting from a sequence length of 128K during the training mid-phase.
The multi-token prediction layer runs the indexer once in the first draft step and reuses it in subsequent steps, eliminating the mismatch between the training/inference KV cache of the previous generation. Combined with rejection sampling and end-to-end TV loss, the accepted length increases from 4.56 to 5.47 — about 20% — across 7 MTP steps.
| Baseline | 4.56 |
| + IndexShare + KVShare | 5.10 |
| + Reject Sampling | 5.29 |
| + End-to-End TV Loss | 5.47 (+20%) |
After more than a few tens of thousands of tokens, the bottleneck is no longer computation, but the KV cache capacity, long context kernel, and CPU overhead. Three fixes: fine-grained memory management and parallelization based on LayerSplit, kernel optimization for context length scaling coordinated with the cache transfer pipeline, and CPU-side cache management, request scheduling, and runtime path optimization. The throughput advantage expands with the growth of context.
Normalized Throughput Advantage
The training stack (slime) combines white-box and black-box rollouts, compact trajectories, and sub-agent workflows. Parallel OPD training merges more than a dozen expert models in about two days, and FP8 KV caching ensures that rollout memory does not exceed the budget.
Long-term RL can induce shortcuts, so suspicious actions are first passed through a rule-based recall filter, and then checked for accuracy by the LLM judge. Confirmed malicious actions are intercepted online and responded with virtual tool results, allowing rollouts to continue rather than be discarded. Based on Critic's PPO, compressible sub-traces are maintained for training on a single rollout.
From 30B Flash level to a million-context flagship — choose a GLM model to view the complete specification table and benchmark data.
One million token context, open-source SOTA for programming and long-term tasks, with stronger agent engineering capabilities.
Previous generation flagship: significantly improved long-term capabilities and engineering-level outputs that can work autonomously for several hours.
Stronger programming capabilities, more reliable multi-step execution, and better complex agent behavior.
Basic model, optimized for long chain and dynamic intelligent agent scenarios.
Enhanced programming capabilities, stable multi-step inference, and improved frontend generation.
300 billion parameters; high efficiency and performance, leading among models of the same scale.
Small-sized GLM model with excellent unit computing power performance.
Real-time speech-to-text recognition, CER as low as 0.0717.
Multi-file functionality beyond a single context window: the model keeps the entire repository state in view, rather than re-reading it every few rounds.
Several hours of experimental cycles, agents need to plan, run, read results, and correct — this is exactly the workload designed to be evaluated by FrontierSWE and PostTrainBench.
Complete call graphs, analyzer outputs, and keeping the kernel and system work on the same track as previous attempts are required.
Long-term reproducibility tracking, unstable testing, and cross-service failures; truncating the context is the reason for losing bugs.
All the above scenarios can be run on your own hardware — MIT weights, transformers · vLLM · SGLang · xLLM · ktransformers.How to run the GLM model locally →
Read the GLM model benchmark tests, and then let the real model start working. The sandbox above is free and requires no information; if you want a more complete AI toolkit, the free package from the partners starts here.