Modelcoding2w ago

GLM-5.2 review

Zhipu's open-weight coding flagship: a 753B mixture-of-experts model with a 1M-token context and MIT weights, claiming to edge past GPT-5.5 on coding benchmarks.

Maker
Zhipu AI (Z.ai)
Launched
Jun 13, 2026
Pricing
open-source
Visit official site
Firstlook

Our verdict

GLM-5.2 is the most credible open-weight challenge to the closed coding leaders yet: MIT weights, a genuinely usable 1M-token context, and benchmark numbers Zhipu says edge past GPT-5.5. The big caveats are that those numbers are the vendor's own, and 'open' here still means a 753B model most people can't self-host. If you want frontier-ish coding you can actually own and audit, it's the one to watch — just verify the benchmarks against your own workload.

First look — our read from the docs and sources below; not yet hands-on tested.

For two years the story of open-weight models was "great, but a generation behind." GLM-5.2, which Zhipu AI released on June 13, 2026 under its international Z.ai brand, is the clearest sign yet that the gap is closing where it matters most — coding — and it does it with MIT-licensed weights you can actually download.

The pitch is specific. It's a 753-billion-parameter mixture-of-experts model with roughly 40B parameters active per token, a genuinely usable 1-million-token context window, and benchmark numbers that Zhipu says nose ahead of the closed leaders: 62.1 on SWE-bench Pro against the ~58.6 it attributes to GPT-5.5, and 99.2 on AIME 2026. If those hold up, it's the first time an openly-licensed model has credibly claimed the coding crown.

That "if" is the whole review. These are vendor-reported numbers — Zhipu marking its own homework. They might be perfectly honest; coding benchmarks are also exactly the kind of metric that gets optimized for. So read 62.1 as a claim worth testing, not a settled result.

Who it's for

Teams who want frontier-ish coding capability they can own — download, audit, fine-tune, run behind their own walls — rather than rent through an API. The 1M-token context is a real differentiator for whole-repo and long-document work, and MIT licensing makes commercial use clean. If you've been waiting for an open model that doesn't feel a generation behind on code, this is the one to actually test.

Who should skip it

If "open weights" was the only reason you were interested, temper it: a 753B MoE is not something most people self-host, so in practice many will use the hosted Coding Plan and the openness becomes more philosophical than practical. And if you need a general-purpose assistant rather than a coding/agentic specialist, or you can't spare the time to verify vendor benchmarks against your own tasks, a settled closed model may serve you better today.

This is a first look from the launch materials, not our own evaluation — so no score from us yet. But on paper, GLM-5.2 is the most interesting open-weight release of the month.

Provider

Providerzhipuglm-5.2· MIT

Specs & key facts

What it isOpen-weight LLM built for coding, reasoning & agentic workloads[src]
Architecture753B mixture-of-experts (~40B active per token)[src]
Context window1M tokens in · 131K out[src]
SWE-bench Pro (vendor-reported)62.1 (vs ~58.6 GPT-5.5, per Zhipu)[src]
AIME 2026 (vendor-reported)99.2[src]
LicenseMIT (commercial use allowed)[src]
Released2026-06-13[src]

Capabilities

CodingYes (primary focus)
Reasoning / mathYes
Agentic / tool useYes
Open weightsYes (MIT)
Self-hostYes (needs serious hardware)
Hosted APIYes (Z.ai)

How to use it

  1. 1Fastest path: use it hosted via the Z.ai GLM Coding Plan or API (OpenAI-compatible).
  2. 2Want to own it? Download the MIT weights — but a 753B MoE needs serious multi-GPU hardware even at ~40B active.
  3. 3It's coding/agentic-tuned, so it shines wired into an agent or IDE assistant with tool access.
  4. 4Lean on the 1M-token context for whole-repo or long-document tasks rather than chunking.

Pricing

Open weights

Free (MIT)

MIT-licensed weights — self-host if you have the hardware for a 753B MoE model. You pay only for the compute.

GLM Coding Plan / API

from ~$18/mo

Hosted access; token pricing reported around $1.40 in / $4.40 out per million tokens.

Open-weight model with an optional hosted plan. Benchmark figures below are vendor-reported (Zhipu's own numbers), not independently verified by us. Verified 2026-06-26.

Pros & cons

Pros

  • MIT-licensed open weights at near-frontier coding scale — rare combination.
  • Usable 1M-token context for whole-repo and long-document work.
  • Vendor benchmarks put it ahead of GPT-5.5 on coding (SWE-bench Pro 62.1).
  • Hosted plan is cheap to try if you don't want to run 753B yourself.

Cons

  • Headline benchmarks are vendor-reported — treat them as claims until independently checked.
  • 'Open' is theoretical for most: 753B MoE needs serious multi-GPU hardware to self-host.
  • Brand-new — tooling, quantizations and real-world reliability are still settling.
  • Coding-tuned, so it's not the obvious pick for general chat or non-code tasks.

Alternatives

FAQ

Sources

Sources

  1. 1.Release date (2026-06-13), 753B MoE (~40B active), 1M context, MIT weights, coding/reasoning/agentic focushttps://www.scmp.com/tech/article/3343239/chinas-zhipu-ai-launches-new-major-model-glm-5-challenge-its-rivalsVerified 2026-06-26
  2. 2.Vendor-reported benchmarks (SWE-bench Pro 62.1, AIME 2026 99.2) + GPT-5.5 comparison; pricinghttps://www.eigent.ai/blog/glm-5-2Verified 2026-06-26
  3. 3.GLM Coding Plan pricing + API accesshttps://z.aiVerified 2026-06-26

More coverage

News & first-looks about this release. Coming soon.
Head-to-head comparisons. Coming soon.