Z.ai ships GLM-5.2: open-source coding model, 1M context

Z.ai released GLM-5.2 on June 22, 2026 — an open-source coding model with a 1M-token context window, MIT license, and benchmarks that put it within striking distance of frontier coding models. It scores 81.0 on Terminal-Bench 2.1, a 19-point leap over GLM-5.1's 62.0, and 62.1 on SWE-bench Pro (up from 58.4). The real headline isn't just the scores — it's the IndexShare architecture, which cuts per-token FLOPs by 2.9x at full context length. This makes long-horizon coding-agent workflows genuinely affordable to run, not just technically possible.

What happened

Z.ai (formerly Zhipu AI), the Chinese AI lab, shipped GLM-5.2 as a direct upgrade to its GLM-5.1 coding model. The model is purpose-built for coding agents that work across large codebases — the kind that reference code, run commands, parse test output, and maintain task history across extended sessions.

The release includes two thinking-effort modes: High for faster responses and Max for harder tasks requiring more compute. Both are available under the MIT license, meaning commercial use, modification, and redistribution are all permitted.

On the benchmark front, the numbers tell a clear story:

Z.ai also reported a best harness result of 82.7 on Terminal-Bench 2.1. For context, Claude Opus 4 scores 85.0 on the same benchmark — GLM-5.2 is close but still trailing (Developer Tech(opens in new tab), 2026).

Early developer reception has been notably positive. Vercel CEO Guillermo Rauch posted on X that he was "impressed" by the model's coding performance. Former Meta, Google DeepMind, and Microsoft executive Matt Velloso wrote that he used GLM-5.2 for a full day and described it as an open model that met his bar for daily use (Developer Tech(opens in new tab), 2026).

Why GLM-5.2 matters

IndexShare changes the economics of long-context inference. It reuses the same indexer across groups of sparse attention layers, cutting per-token FLOPs by 2.9x at 1M-token context. Z.ai also improved its multi-token prediction layer, increasing speculative decoding acceptance length by up to 20%. Together, these aren't paper optimizations — they directly lower the cost of running coding agents that process entire repositories.

MIT license with self-hosting removes the vendor gate. GLM-5.2 runs on vLLM, SGLang, Hugging Face Transformers, Docker Model Runner, and KTransformers — no proprietary API required. Documentation also lists support for Ascend NPU platforms via vLLM-Ascend, xLLM, and SGLang. For enterprise teams with GPU infrastructure, this means full control over deployment, data handling, and cost.

The geopolitical catch is real. Z.ai is a Chinese lab, and export controls remain a moving target. The model is technically strong — but organizational policy, not technical merit, will be the deciding factor for many teams in regulated environments.

What GLM-5.2 changes for you

• If you self-host open-source models: GLM-5.2 is the strongest MIT-licensed coding model benchmarked to date for long-horizon tasks. Test it head-to-head against your current coding-agent model on your longest workflows — the IndexShare architecture means it may actually be cheaper to run at full context.
• If you use closed APIs: The benchmark proximity to Claude Opus 4.8 makes GLM-5.2 worth watching. It won't replace a hosted API for every team, but it narrows the gap between open-source and frontier coding models more than any MIT-licensed release before it.
• If your org has geopolitical constraints: Wait. The model is good, but policy risk is a genuine blocker. Track whether Z.ai's models gain broader enterprise adoption before committing production infrastructure.

FAQ

What makes IndexShare different from standard attention? Standard attention computes per-token independently, which gets expensive at long context. IndexShare reuses a shared indexer across groups of sparse attention layers, cutting FLOPs 2.9x at 1M tokens. Think of it as amortizing the cost of deciding which tokens to attend to.

How does GLM-5.2 compare to Llama 4 Maverick for coding? GLM-5.2 is purpose-built for coding agents and scored 81.0 on Terminal-Bench 2.1 — Llama 4 Maverick's published benchmarks don't target this class. Llama 4 Maverick remains the safer open-source bet for general-purpose workloads and teams that value Meta's broader ecosystem. GLM-5.2 wins on raw coding-agent numbers but carries geopolitical risk that Llama doesn't.

Can I fine-tune GLM-5.2? Yes. The MIT license permits modification, and the model runs on standard inference frameworks like vLLM and Hugging Face Transformers. Z.ai hasn't published fine-tuning recipes yet, but the open weights mean the community can — and likely will.