The full refusal-ablated GLM‑5.2 research base — every expert intact, NVFP4-quantized for the reference serve. MTP speculative decode at 2.1×, up to 1M context — for the most demanding red-team and evaluation work.
GLM‑5.2 NVFP4 is the reference research serve — the full refusal-ablated GLM‑5.2 with all 256 experts intact, quantized weight-only to NVFP4 so it fits and flies on Blackwell. The W4A16 multi-token-prediction fix delivers ~56 tok/s (2.1×) single-stream, with 512K context out of the box and headroom to 1M. Zero true refusals across a 450-prompt benchmark — a clean, unclipped control for evaluating safety systems and adversarial robustness when you have the eight cards.
| Base | GLM-5.2 (refusal-ablated) |
| Architecture | MoE · DSA sparse attn |
| Experts | 256 / 256 (full) |
| Quantization | NVFP4 · W4A16 · group-16 |
| Kept at BF16 | attention · lm_head · norms |
| Context | 512K (→ 1M, YaRN) |
| Throughput | ~56 tok/s · MTP 2.1× |
| Refusal benchmark | 0 / 450 adversarial |
| Weights | ~420 GB |
| Hardware | 8× RTX PRO 6000 (Blackwell) |
| License | commercial + research |
Eight-card serves where you want the full model with nothing pruned for your evaluations.
Benchmark guardrails and classifiers against an unclipped reference baseline.
Long-transcript analysis, whole-repo agents, and extended adversarial sessions.
GLM-5.2 REAP runs the same research base on half the GPUs; BF16 gives you full precision to fine-tune your own evaluation and detector models.