The full-precision reference weights — the refusal-ablated GLM‑5.2 exactly as delivered. Fine-tune it, distill from it, or roll your own quants for evaluation and detector work.
GLM‑5.2 BF16 is the unquantized, full-precision checkpoint — the same refusal-ablated GLM‑5.2 base behind the NVFP4 builds, delivered raw. This is the build for teams that need to fine-tune a domain-specific red-team model, distill a smaller evaluation model, or produce their own quantizations (GGUF, AWQ, custom NVFP4). Full MoE, DSA sparse attention, native long context, a clean 0-refusal baseline to build on.
| Base | GLM-5.2 (refusal-ablated) |
| Architecture | MoE · DSA sparse attn |
| Experts | 256 / 256 (full) |
| Precision | bfloat16 (unquantized) |
| Context | 512K (→ 1M, YaRN) |
| Refusal benchmark | 0 / 450 adversarial |
| Weights | ~1.5 TB |
| Format | safetensors · HF Transformers |
| Best for | fine-tune · distill · quantize |
| License | commercial + research |
Adapt the base into a domain-specific red-team or evaluation model from a clean full-precision start.
Produce your own GGUF / AWQ / NVFP4 research targets from the source weights.
Distillation, interpretability, and safety evaluation on the unquantized model.
The NVFP4 builds are ready to deploy — full on eight cards, or REAP-pruned on four.