GLM-5.3 BF16 safetensors (full precision)

Z.ai's GLM-5.3, released August 2026 — the same 753B-parameter mixture-of-experts architecture as GLM-5.2 (glm_moe_dsa: MoE transformer with DSA sparse attention, 1M-token context, English/Chinese), with a new post-training round aimed at coding and agentic engineering: Z.ai reports roughly 50% better coding results than GLM-5.2 and state-of-the-art scores on Terminal Bench 3.0 and vulnerability-discovery benchmarks. Supports a reasoning_effort control (low / high / max).

License note: the GLM-5.3 License is permissive for research and commercial use; its only restriction is a Model-as-a-Service clause that applies to organisations above US$10B annual revenue. Read the license before building a hosted service on it.

This torrent mirrors the complete BF16 full-precision release weights (292 files, 32 MiB pieces) from the official Hugging Face GLM-5.3-BF16 repository — the unquantised weights rather than the FP8 main repo. Verify your download against the SHA-256 checksums on the model card.

Swarm verified via peer metadata fetch on 2026-08-29; size and file count are taken from the torrent metadata itself. Created and permanently seeded by drforbin.ai — this is the site's own release of the official Hugging Face weights.