What you need to know
- Open-weight, MIT licence. The weights are downloadable from Hugging Face and ModelScope under permissive MIT terms — download, fine-tune, ship commercially, self-host. It is open-weight, not open-source: the training data and pipeline are not published.
- Frontier-adjacent on coding. 81.0 on Terminal-Bench 2.1, 62.1 on SWE-bench Pro, and within ~1% of Opus 4.8 on FrontierSWE — while beating GPT-5.5 on multiple long-horizon coding tasks.
- Roughly a sixth of the cost. Z.ai prices the metered API at $1.40 per million input tokens and $4.40 per million output tokens, or a flat GLM Coding Plan from about $18 a month.
- The catch is data residency. The hosted API runs through Z.ai's China-based infrastructure. For regulated or personal data, self-host the MIT weights instead so nothing crosses a border you cannot account for.
Treat GLM-5.2 as two products, not one. The hosted API is the cheapest way to prototype and to run public, non-sensitive workloads. The downloadable weights are the way to put a frontier-adjacent coder behind your own firewall for anything touching customer or employee data. Most teams will use both — API for speed, self-host for compliance.
What shipped, and what the MIT licence actually buys you
GLM-5.2 is the latest flagship from Zhipu AI, shipping under the company's international Z.ai brand. It is a Mixture-of-Experts model: around 753 billion total parameters, but only about 40 billion active per token, which is what keeps inference tractable. It carries a 1M-token context window and can emit up to 131,072 tokens in a single response — generous enough for whole-repository reasoning and long agentic runs.
The part that matters most is the licence. The weights ship under MIT — one of the most permissive licences in software. You can download them, fine-tune on your own data, deploy commercially, and run them on hardware you control, with no royalty and no usage gate. That is a meaningfully different proposition from a closed API where the model lives on someone else's servers and your only lever is the price list.
One precision worth keeping: this is open-weight, not open-source. Z.ai has released the trained weights, not the full training corpus or pipeline, so you cannot reproduce the model from scratch or fully audit how it was trained. For most builders that distinction is academic — you want a capable model you can host — but it matters if your compliance posture requires provenance over the training data itself.
The benchmarks, in context
Open-weight coders have been closing on the closed frontier all year, and GLM-5.2 is the clearest evidence yet that the gap is now small enough to ignore for a lot of real work. The numbers, with the exact benchmark variants:
| Benchmark | GLM-5.2 | Reference point |
|---|---|---|
| Terminal-Bench 2.1 | 81.0 | Long-horizon agentic shell tasks |
| SWE-bench Pro | 62.1 | Harder, real-world PR resolution |
| FrontierSWE | ~1% behind Opus 4.8 | Trails Anthropic's flagship by roughly one point |
| vs GPT-5.5 (long-horizon coding) | Wins on several | At roughly 1/6th the cost per task |
Read these the way you would read any vendor-reported figure: as a strong signal to run your own evaluation, not as gospel. Terminal-Bench 2.1 and SWE-bench Pro both stress long-horizon, multi-step agentic behaviour — exactly the workload where cheaper models historically fell apart — and an 81.0 and 62.1 respectively put GLM-5.2 firmly in the conversation with the closed flagships. Trailing Opus 4.8 by about a single point on FrontierSWE is, for an MIT-licensed model you can run on your own GPUs, a remarkable place to be. If you want a primer on reading these leaderboards critically before you trust a headline number, our explainer on how to read the June 2026 SWE-bench leaderboard is worth ten minutes.
The cost story — the one-sixth claim, examined
Cost is where this stops being an academic exercise. Z.ai offers two ways to pay:
- Metered API: $1.40 per million input tokens and $4.40 per million output tokens.
- GLM Coding Plan: a flat subscription from roughly $18 a month for steady coding use.
Against GPT-5.5, that works out to roughly a sixth of the per-task cost on the long-horizon coding workloads Z.ai benchmarked — and Z.ai is claiming a win on capability there too, not a trade-off. For a small team in Bengaluru or Manchester running a coding agent across a real codebase, the difference between a sixth and full price is the difference between an experiment that pays for itself and one that quietly burns the runway.
| Option | Input / output ($ per MTok) | Best for |
|---|---|---|
| GLM Coding Plan | Flat from ~$18/month | Steady, predictable coding sessions |
| GLM-5.2 metered API | $1.40 in / $4.40 out | Bursty or variable workloads |
| Self-host MIT weights | Your GPU cost only | Sensitive data, high steady volume |
The self-host column has no per-token line because your only cost is the GPU you rent or own. At high, steady volume that flips the economics: once you are saturating a node, paying for the hardware can beat paying per token — and you get data residency thrown in. We worked through where that break-even sits in our guide on choosing between self-hosting and an API for LLM inference; the short version is that volume and data-sensitivity are the two variables that decide it.
Hosted API versus self-host — and the data-residency caveat
Here is the nuance that every Indian and UK builder needs to sit with before wiring GLM-5.2 into production.
The hosted Z.ai API runs on China-based infrastructure, which carries cross-border data-residency considerations. Sending customer records, employee data, or anything personal through it means that data may be processed outside India, the UK or the EU. Under India's DPDP Act and under UK and EU data-protection rules, that can put you on the wrong side of your obligations. For any sensitive or regulated workload, self-host the MIT weights so the data never leaves a jurisdiction you control.
This is not a reason to dismiss the model — it is a reason to use it deliberately. The decision tree is simple:
- Non-sensitive, public, or synthetic data? Use the hosted API. It is the cheapest, fastest path and there is nothing to lose to a border crossing.
- Personal, financial, health, or otherwise regulated data? Self-host the weights on your own hardware or a regional GPU provider, so processing stays inside India, the UK or the EU.
- Not sure yet? Default to self-host for anything customer-facing. It is far cheaper to start compliant than to retrofit residency after an audit.
Because the weights are MIT-licensed, the self-host path is genuinely open to you — this is exactly the scenario open weights exist for. A practical, step-by-step route to a production deployment on vLLM is in our guide to self-hosting an open-weight LLM in production; it covers the serving stack, quantisation, and the throughput tuning that decides whether your GPU bill makes sense.
Every article here is written by a Verified Builder. Want your name on the next one?
AI Tech Connect lists AI engineers, founders and researchers across India and the UK — and the people hiring browse it to find them. Adding your profile is free.
Become a Verified Builder →Where GLM-5.2 sits in the open-weight wave
GLM-5.2 does not arrive in a vacuum. June 2026 has been the month the open-weight coding wave crested — Kimi K2.x, the broader GLM line, and MiniMax M3 all landing in quick succession, each pushing capability up and price down. We mapped the whole wave in our piece on the open-weight coding models from Kimi, GLM and MiniMax, and looked closely at MiniMax M3's open-weight 1M-context release, which sits in much the same competitive slot as GLM-5.2.
What is striking is how international the front line has become. GLM-5.2 comes from a Chinese lab via an international brand; Europe's strongest open contender, Mistral Large 3, topped an open leaderboard as the EU-jurisdiction option; and India's own sovereign models are coming up fast. For a builder, the practical upshot is choice: you are no longer picking between a closed frontier model and a weak open one. You are picking between several frontier-adjacent open-weight options, each with different licence terms and different residency stories. GLM-5.2's MIT licence is, on the licence axis, about as friendly as it gets.
What Indian and UK builders should actually do
Concrete guidance, by situation:
- Solo builder or early-stage startup, non-sensitive code: Start on the hosted API or the GLM Coding Plan. At a sixth of GPT-5.5's cost for comparable long-horizon coding, it is the most capital-efficient way to ship right now.
- Team handling customer or personal data: Self-host the MIT weights from day one. In India, IndiaAI's subsidised GPU-hours make a multi-GPU node affordable; in the UK, regional cloud regions keep processing inside UK or EU jurisdiction. Either way, DPDP and UK and EU rules are satisfied because the data never leaves.
- Already on a closed coding model: Run a side-by-side evaluation on your own tasks before switching. The benchmarks are encouraging, but your codebase and your prompts are the only test that counts. Route the cost-sensitive, high-volume work to GLM-5.2 and keep your closed model for the hardest edge cases if it still wins on your evals.
- Building agentic workflows: The 1M-token context and 131,072-token output ceiling make GLM-5.2 a credible base for long-running agents and whole-repo refactors — the workloads Terminal-Bench 2.1 is designed to probe.
The pragmatic default for most teams is split routing plus self-host for sensitive data: send cheap, high-volume, non-sensitive coding tasks to GLM-5.2's API, run anything touching personal data on self-hosted weights, and reserve a closed flagship only for the small set of tasks where your own evals show it still wins. You get frontier-adjacent quality, a sixth of the cost, and a clean data-residency story all at once.
The bottom line: GLM-5.2 is not a curiosity. It is a genuinely usable, self-hostable, frontier-adjacent coding model with the most permissive licence in its class. For builders in India and the UK who care about both their cloud bill and their data-protection obligations, that combination is rare enough to be worth a serious evaluation this quarter.
The weights are on Hugging Face and ModelScope; pricing and the GLM Coding Plan are on z.ai.