What Is GLM-5.2? Z.ai's 1M-Context Open-Weight Flagship, Two Versions Later

GLM-5.2 is an open-weight, text-only large language model from Z.ai, the Beijing lab formerly known as Zhipu AI, released on 16 June 2026 under the MIT licence. It is a 753-billion-parameter mixture-of-experts model with a one-million-token context window, and for roughly two months it was the strongest open-weight coding model anyone could download. It no longer holds that position, because Z.ai has since shipped two successors on top of it — which is the part most pages about GLM-5.2 leave out, and the part that decides whether you should still care.

Forty-three articles in this blog's archive mention GLM-5.2. Until now, not one of them was about it: the model appears as the opponent in comparison after comparison, the fixed reference point that newer models get measured against. That is a strange role for a model that its own maker has already moved past, and it is the reason this page exists — a straight description of what GLM-5.2 actually is, what it costs, what it is good at, and who should still be picking it.

Where GLM-5.2 sits in Z.ai's own line

Z.ai publishes a dated release log, and it is the cleanest source for this. Read it in order and the shape of the family is obvious:

• GLM-5 — 12 February 2026. The first GLM-5, aimed at complex systems engineering and long-range agent tasks.
• GLM-5.1 — 7 April 2026. Long-horizon work, able to run solo for up to eight hours in a single run.
• GLM-5.2 — 16 June 2026. The 1M-context flagship for long-horizon tasks; Z.ai's log describes it as "1M lossless context, significantly improving long-horizon task capabilities".
• GLM-5.3 — 18 August 2026. Same base model as GLM-5.2, with the gains coming from post-training. Z.ai reports a 50% gain over GLM-5.2 on its in-house Code Bench and open-source state of the art on Terminal Bench 3.0.
• GLM-5.3-Flash — 26 August 2026. A different, smaller model built on a newly trained base (320B total, 18B active), the first natively multimodal GLM-5.

The important line is the GLM-5.3 entry. GLM-5.3 is not a bigger GLM-5.2 — it is the same base weights pushed further with post-training. That makes GLM-5.2 the last model in the line whose capability came from the architecture, and it makes GLM-5.3 the current text flagship. If you are choosing a Z.ai model today on quality alone, GLM-5.3 is the answer and GLM-5.2 is not.

GLM-5.3-Flash is a separate branch rather than a cheaper GLM-5.3: a smaller, natively multimodal model (text, image and video) priced an order of magnitude below the text flagships. If you need vision, GLM-5.2 is not the model you want in either case.

The spec sheet

• Parameters — 753B total, per the model card at zai-org/GLM-5.2-FP8. Z.ai does not state the active-parameter count on that card; third-party listings put it near 40B per token, and we could not confirm that figure from a primary source.
• Context window — 1M tokens, described on the card as "a solid 1M-token context that stably sustains long-horizon work".
• Maximum output — 128K tokens.
• Modality — text in, text out. No image input, no audio. Our own catalogue entry for the model states it "supports only text input".
• Licence — MIT, with no regional limits. Weights are published through the zai-org organisation on Hugging Face in BF16 and FP8 variants.
• Reasoning control — a hybrid thinking mode driven by a reasoning_effort parameter with high and max settings, defaulting to max. Native tool calling and structured output are supported.
• Architecture — IndexShare, which reuses one lightweight indexer across every four sparse-attention layers and, per the model card, cuts per-token FLOPs by 2.9× at full context; plus an improved multi-token-prediction layer that raises speculative-decoding acceptance length by up to 20%.
• Training hardware — press coverage of the launch states the model was trained on Huawei Ascend accelerators with no Nvidia hardware. That is a claim from news reporting, not from the model card, and we are passing it on as such.

What GLM-5.2 is measurably good at

Almost everything GLM-5.2 is good at is a coding or agentic measurement, and the numbers below are vendor-reported — they come from the benchmark table on Z.ai's own model card, not from an independent reproduction.

• Terminal Bench 2.1 (Terminus-2) — 81.0, against 63.5 for GLM-5.1. That is the single largest generational jump in the table.
• SWE-bench Pro — 62.1, against 58.4 for GLM-5.1.
• FrontierSWE (Dominance) — 74.4, against 30.5 for GLM-5.1 and 72.6 for GPT-5.5.
• AIME 2026 — 99.2. GPQA-Diamond — 91.2. Both comfortably competitive with the closed frontier on Z.ai's own table.
• MCP-Atlas (public set) — 76.8, essentially level with Claude Opus 4.8's 77.8 on the same table.
• Long-context recall — 78.3, which is the number that matters most for the 1M-token window. The window is only useful if recall holds at depth, and Z.ai's own figure says it mostly does.

The independent picture is thinner but real. Artificial Analysis carries GLM-5.2 at an Intelligence Index score of 34 on its Intelligence Index v4.3.2, placing it 11th of 114 models in its class. Two caveats attach to that number, and Artificial Analysis states both itself: the model is marked deprecated on their page, and they "only continue performance benchmarking for the default 10k input token workload" — results for other workloads are historical and no longer updated. Our own catalogue entry for the model carries an older evaluation snapshot from 25 June 2026 with an Intelligence score of 33.7, which is not the same index revision as the live figure and should not be read as a change in the model.

What it is measurably bad at

Three things, and they are worth being blunt about.

• It is text-only. No image input, no audio input. GLM-5.3-Flash is the multimodal model in this family, and if your workload touches screenshots, PDFs or video, GLM-5.2 is the wrong branch entirely.
• It loses badly on the longest-horizon and hardest software-engineering measurements. On SWE-Marathon it scores 13.0 against Claude Opus 4.8's 26.0. On DeepSWE it scores 46.2 against 58 for Claude Opus 4.8 and 70 for GPT-5.5. On NL2Repo it scores 48.9 against Opus 4.8's 69.7. These are all vendor-reported figures from the same table that shows GLM-5.2 winning elsewhere, which is what makes them credible: Z.ai published them next to its own wins.
• It has been superseded on the measurements that made it famous. GLM-5.3 uses the same base and beats it on Z.ai's own Code Bench by 50%, on Terminal Bench 3.0 and on Agents' Last Exam. GLM-5.2 is also flagged deprecated at Artificial Analysis.

One figure we could not obtain: we could not confirm a current independent throughput or latency measurement for GLM-5.2 from a source we were able to open, so this page states none rather than repeating a number we could not check.

Price and where to run it

Z.ai's own rate card, read today, lists GLM-5.2 at:

• Input — $1.40 per million tokens
• Cached input — $0.26 per million tokens
• Output — $4.40 per million tokens

Cached-input storage is listed as limited-time free. Note that GLM-5.3 carries exactly the same three rates, so the newer model is not more expensive on Z.ai's list — which removes the usual reason to stay on the older one.

On availability: the model is served through Z.ai's own OpenAI-compatible endpoint at api.z.ai/api/paas/v4/chat/completions, with official SDKs for Python and Java, and the weights are downloadable from Hugging Face for self-hosting under MIT. Beyond that it is available through a number of third-party inference platforms. We route it: OrcaRouter carries GLM-5.2 on its model page at the Z.ai provider rate with zero markup, so the $1.40 / $4.40 above is what you pay through us rather than a marked-up copy of it. The same key also reaches the rest of the catalogue, which matters here for a specific reason — see below.

Who should pick GLM-5.2, and who should pick something else

Pick GLM-5.2 if you are self-hosting and the 1M-token window is the requirement rather than the benchmark scores. The MIT licence with no regional restrictions, the published FP8 weights and the IndexShare attention design make it a genuinely practical thing to run at long context, and the recall figure at depth is the number that justifies the window. It is also a reasonable pick if you have a pipeline already tuned to its reasoning_effort behaviour and the cost of re-validating prompts against GLM-5.3 exceeds the gain.

Pick GLM-5.3 instead if you are buying tokens rather than weights and quality is the only axis. It is the same base model, better post-trained, at identical list prices. There is no pricing argument for the older model on Z.ai's own rate card.

Pick GLM-5.3-Flash instead if you need vision, or if you need the cheapest capable option: it handles text, image and video and is priced far below both text flagships.

Pick something else entirely if your work is at the hard end of long-horizon software engineering. On Z.ai's own published table, Claude Opus 4.8 beats GLM-5.2 on SWE-Marathon, DeepSWE, NL2Repo, SWE-bench Pro, ProgramBench and Tool-Decathlon; GPT-5.5 beats it on DeepSWE and ProgramBench. GLM-5.2's wins are real but they are concentrated in terminal-style agentic coding and reasoning, not in the marathon tasks. If your evaluation looks like a two-hour agent run against a large unfamiliar repository, the open-weight price advantage does not close that gap on these numbers.

The practical summary

GLM-5.2 is a 753B-parameter, MIT-licensed, 1M-context, text-only MoE model released on 16 June 2026 at $1.40 / $4.40 per million tokens, with vendor-reported Terminal Bench 2.1 of 81.0, SWE-bench Pro of 62.1 and an independent Artificial Analysis Intelligence Index of 34. It is the last GLM flagship whose capability came from architecture rather than post-training, it has been superseded twice by its own maker, and it is marked deprecated on the independent index that scored it.

None of that makes it a bad model. It makes it a settled one: the interesting question about GLM-5.2 in September 2026 is not whether it is good, but whether the specific thing you need — a long-context, self-hostable, MIT-licensed coding model — is worth more to you than the three points GLM-5.3 adds on top of the same weights. For most buyers of tokens the answer is no. For anyone downloading weights, it still is.

If you want to test that for yourself without committing a production path to it, the routing layer is the cheap way to do it: point the same request at GLM-5.2 and GLM-5.3 through one endpoint, compare on your own prompts, and let automatic failover handle the case where the older model's endpoint is the one that goes down.

Call GLM-5.2 through OrcaRouter at the Z.ai provider rate, zero markup. https://www.orcarouter.ai/models/z-ai/glm-5.2 One API key reaches GLM-5.2, GLM-5.3 and 200+ other models, with automatic failover if an endpoint goes down.