GLM-5.2 Ships 753B Parameters Under a Plain MIT License

Z.ai released GLM-5.2 as a 753-billion-parameter mixture-of-experts model with a one-million-token context window, under an unmodified MIT license with no regional restrictions. That combination is unusual: most frontier-tier open-weight releases ship under bespoke licenses with use restrictions or naming obligations attached. The technical headline is IndexShare, a mechanism that reuses indexers across sparse attention layers and, per the model card, cuts per-token FLOPs by 2.9× at full context — attacking the specific overhead that normally erodes sparse attention’s theoretical savings at long context. Support is already broad across vLLM, SGLang, Transformers, KTransformers and Unsloth. For teams whose blocker on open weights is legal review rather than capability, MIT at this scale removes the blocker outright, and that may prove a more effective distribution strategy than any benchmark score.