GLM-5.2: Z.ai's MIT-licensed MoE with a 1M context and a sparse-attention fix

Z.ai documents a 1-million-token context window, up from roughly 200K on GLM-5.1, with a reported maximum output of about 128K tokens. The company's framing is that the context is usable for long-horizon engineering work rather than merely advertised, which is the more useful claim to test: a context length that degrades in quality partway through is not the same capability as one that holds up.

Third-party reporting describes the model as positioned between two closed frontier models on coding capability under similar token budgets, and Z.ai's own tables report improvements on Terminal-Bench 2.1 and SWE-bench Pro over the previous generation. Those are vendor and third-party measurements on their own configurations. Whether the context is reliable for your task is a measurement only you can make, on your own repository and harness.

Z.ai describes IndexShare as reusing the same lightweight indexer across every fourth sparse-attention layer, which it says reduces per-token compute substantially at a 1M context length, and describes an improved multi-token-prediction layer that increases acceptance length for speculative decoding. Both are efficiency claims about serving long context, and both are the kind of claim that is checkable: run the model and measure.

The architecture also constrains which runtimes can serve the model. Z.ai lists support for major inference frameworks, and third-party reporting records roughly 753 billion total parameters with about 40 billion active per token. A sparse model of that size is not a single-GPU proposition, so the practical question is which serving stack and hardware your team already operates.

Z.ai publishes GLM-5.2 weights under the MIT licence and distributes them on major model hubs. MIT permits commercial use, modification, and redistribution with attribution, which makes self-hosting a genuine commercial option rather than a research exception. It also means a modified model you train can be deployed and sold without a vendor agreement.

The licence is the easy part. Read it for the exact checkpoint you intend to run, record the version and the date you checked it, and treat the hosted endpoint's terms as a separate document — a model's weight licence and a provider's API terms are not the same thing, and neither substitutes for the other.

GLM-5.2 is reachable at zero cost on at least one gateway. The published commercial price is roughly a dollar and forty cents per million input tokens, twenty-six cents cached, and four dollars forty per million output. That gap between the free tier and the commercial price is the clearest statement of what the free endpoint costs someone else.

For evaluation, prototyping, and low-volume internal work, free access is genuinely useful. For anything that must keep running, the MIT licence gives you a second route — run it yourself — which converts a subsidy you do not control into a cost you can forecast, subject to the hardware question above. Design that fallback before the dependency, not after an endpoint stops responding.