GLM-5.2 is the flagship open-weight GLM AI model: a solid 1M-token context, Non-Thinking / High / Max effort levels, and reported scores of 81.0 on Terminal-Bench 2.1 and 74.4 on FrontierSWE. Compare every GLM model below — or try the free AI playground first.
No sign-up, no account, no credits. Type a prompt and get a streamed answer.
Response will stream here...
This playground runs a free open-source LLM via OpenRouter. It is not GLM-5.2 and does not use GLM weights — it's here so you can test a live model while you read the published GLM benchmark data below.
Four capabilities define the current GLM model generation — all of them reported by the model authors and reproducible from open weights.
A 1M-token context that stays usable, not just accepted: GLM-5.2 was trained on long-context coding-agent trajectories covering large-scale implementation, automated research and complex debugging.
Non-Thinking, High and Max let you trade latency for capability per task — roughly 63% at ~35K output tokens up to 74% at ~83K on agentic coding evals.
One lightweight indexer shared across every four sparse-attention layers cuts per-token FLOPs by 2.9× at 1M context length, with no quality trade at long range.
An open-source GLM model under MIT: weights on HuggingFace and ModelScope, no regional limits, and local serving through transformers, vLLM, SGLang, xLLM and ktransformers.
All figures below are results published by the model authors for GLM-5.2 and its comparison set. glmmodel.com reports them; it does not run these evaluations.
Dominance, tasks up to 20 hours
Up to 10 hours on a single H100
Ultra-long-horizon SWE, up to 10 hours
Across all three long-horizon benchmarks, GLM-5.2 is the highest-ranked open-source model — evidence that its 1M context translates into delivered work, not just accepted tokens. It trails Opus 4.8 by one point on FrontierSWE, beats GPT-5.5 by one, and beats the previous-generation Opus 4.7 by eleven.
Published GLM model benchmark results, 17 evaluations across 8 models.
| Benchmark | 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 |
|---|---|---|---|---|---|---|---|---|
| Reasoning | ||||||||
| HLE | 40.5 | 31.0 | 41.4 | 37.0 | 37.7 | 49.8* | 41.4* | 45.0 |
| HLE w/ 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 Nov. 2025 | 94.4 | 94.0 | 95.0 | 84.4 | 94.4 | 96.5 | 96.5 | 94.8 |
| HMMT Feb. 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 |
| Coding | ||||||||
| 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 harness) | 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 |
| Agentic | ||||||||
| MCP-Atlas (public set) | 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 |
* score on the full set.
A GLM model does not have one speed. Effort level control lets you spend more computation only on the tasks that need it.
~35K avg tokens
Fast edits, boilerplate, routine refactors where latency matters more than depth.
~43K avg tokens
The default for most agentic coding: near-Max quality at roughly half the token spend.
~83K avg tokens
Hard, long-horizon problems — allocate extra compute when a task genuinely warrants it.
Averaged over Terminal-Bench 2.1, DeepSWE and SWE-Atlas QnA, evaluated on Claude Code 2.1.167. Reported by the model authors. At comparable token budgets GLM-5.2 sits between Claude Opus 4.7 and Opus 4.8.
Raising a context ceiling from 200K to 1M tokens is an engineering problem, not a config flag. Three changes do the work.
Every four transformer layers share one lightweight indexer. It sits on the first of the four, and its top-k indices are reused by the other three — removing the indexer dot-product and top-k work in 3 of every 4 layers. Result: 2.9× lower per-token FLOPs at 1M context length. Trained with IndexShare from mid-training at 128K sequence length.
The multi-token-prediction layer runs its indexer once on the first draft step and reuses the indices for later steps, removing the training/inference KV-cache mismatch that limited the previous generation. Combined with rejection sampling and an end-to-end TV loss, acceptance length rises from 4.56 to 5.47 — about 20% — across 7 MTP steps.
| Baseline | 4.56 |
| + IndexShare + KVShare | 5.10 |
| + Rejection Sampling | 5.29 |
| + End-to-end TV Loss | 5.47 (+20%) |
Past a few hundred thousand tokens the bottleneck stops being compute and becomes KV-cache capacity, long-context kernels and CPU overhead. Three fixes: LayerSplit-based fine-grained memory management and parallelism, context-length-scaling kernel work coordinated with the cache-transfer pipeline, and CPU-side cache management, request scheduling and runtime path tuning. The throughput advantage widens as context grows.
Normalized Throughput Advantage
The training stack (slime) combines white-box and black-box rollout, compact trajectories and sub-agent workflows. Parallel OPD training merged more than ten expert models in roughly two days, with FP8 KV-cache keeping rollout memory in budget.
Long-horizon RL invites shortcuts, so suspect actions pass a rule-based recall filter, then an LLM-judge precision check. Confirmed hacks are blocked online and answered with a dummy tool result, letting the rollout continue instead of being thrown away. Critic-based PPO on individual rollouts keeps compaction sub-traces trainable.
From the 30B Flash tier to the 1M-context flagship — pick a GLM model and see its full spec sheet and benchmark column.
1M-token context, open-source SOTA on coding and long-horizon tasks, stronger agentic engineering.
The previous flagship: large long-horizon gains and multi-hour autonomous work at engineering-grade output.
Stronger coding, more reliable multi-step execution and better complex agent behaviour.
Base model tuned for long-chain, dynamic agent scenarios.
Enhanced programming, stable multi-step reasoning and improved front-end generation.
30B parameters; efficient and high-performance, ahead of similar-scale open models.
Small-footprint GLM model with strong performance per unit of compute.
Real-time speech-to-text recognition at 0.0717 CER.
Multi-file features that outlive a single context window: the model keeps the whole repo state in view instead of re-reading it every few turns.
Hours-long experiment loops where an agent must plan, run, read results and revise — the workload FrontierSWE and PostTrainBench were built to score.
Kernel and systems work that needs the full call graph, profiler output and prior attempts held together in one trajectory.
Long reproduction traces, flaky tests and cross-service failures, where truncating the context is what loses the bug.
Run any of it on your own hardware — MIT weights, transformers · vLLM · SGLang · xLLM · ktransformers. How to run a GLM model locally →
Read the GLM model benchmarks, then put a real model to work. The playground above is free and needs nothing from you; if you want a fuller AI toolkit, our partner's free tier starts here.