The refusal-ablated GLM‑5.2, REAP‑pruned to run on four Blackwell cards instead of eight — a full-fidelity research baseline for red-teaming and robustness evaluation, MTP-accelerated with 512K context.
GLM‑5.2 REAP takes our refusal-ablated GLM‑5.2 and applies REAP — router-weighted expert-activation pruning — to drop roughly 31% of the experts (a non-uniform 176-of-256 average) while holding quality near the unpruned model. The result is a ~302 GB NVFP4 checkpoint that serves on 4× RTX PRO 6000 where the full model needs eight — a controlled baseline for red-team exercises, defense evaluation, and robustness benchmarking, run entirely in your own environment. MTP speculative decode is baked in for ~52 tok/s, with a native 512K context window.
| Base | GLM-5.2 (refusal-ablated) |
| Architecture | MoE · DSA sparse attn |
| Experts | ~176 / 256 (non-uniform) |
| Pruning | REAP-NU176, near-lossless |
| Quantization | NVFP4 · W4A16 · group-16 |
| Kept at BF16 | attention · lm_head · norms |
| Context | 512K (native 1M arch) |
| Throughput | ~52 tok/s · MTP on |
| Refusal benchmark | 0 / 450 adversarial |
| Weights | ~302 GB |
| Hardware | 4× RTX PRO 6000 (Blackwell) |
| License | commercial + research |
Blackwell iron but not eight cards — run the full-fidelity research baseline that fits, in-house.
Hold whole attack corpora and long transcripts in a 512K context for single-pass evaluation.
MTP throughput for high-volume benchmark and defense-testing runs — the best tok/s-per-dollar in the lineup.
GLM-5.2 NVFP4 keeps all 256 experts for the reference serve; BF16 gives you full precision to fine-tune evaluation and detector models.