GLM-5.2 · NVFP4 · REAP-NU176

GLM‑5.2 REAP

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.

Red-team baseline · 0/450 Runs on 4 cards 512K context MTP · ~52 tok/s
One-time · commercial + research
$399
Buy GLM‑5.2 REAPBitcoin — $359 · save 10%
Secure checkout via Stripe
Gated research access approved after purchase
Overview

The full research base on half the GPUs.

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.

Specifications
BaseGLM-5.2 (refusal-ablated)
ArchitectureMoE · DSA sparse attn
Experts~176 / 256 (non-uniform)
PruningREAP-NU176, near-lossless
QuantizationNVFP4 · W4A16 · group-16
Kept at BF16attention · lm_head · norms
Context512K (native 1M arch)
Throughput~52 tok/s · MTP on
Refusal benchmark0 / 450 adversarial
Weights~302 GB
Hardware4× RTX PRO 6000 (Blackwell)
Licensecommercial + research
What's included
  • NVFP4 weights — gated Hugging Face research download
  • Deploy kit: prebuilt serve image recipe, launch scripts, the non-uniform expert patch, and the W4A16 MTP patch
  • Think / no-think chat templates
  • Build & serving guide — from `docker run` to first token
  • Commercial + research license
Quickstart

Stand it up with the included kit.

# from the deploy kit — 4× RTX PRO 6000
MTP=1 MAX_MODEL_LEN=524288 bash serve_glm52_reap.sh

# OpenAI-compatible endpoint on :8000 — point your eval harness at it
curl localhost:8000/v1/chat/completions -d '{"model":"glm-5.2-reap","messages":[...]}'
Who it's for

AI red-team labs

Blackwell iron but not eight cards — run the full-fidelity research baseline that fits, in-house.

Robustness researchers

Hold whole attack corpora and long transcripts in a 512K context for single-pass evaluation.

Evaluation pipelines

MTP throughput for high-volume benchmark and defense-testing runs — the best tok/s-per-dollar in the lineup.

Other builds

Need max fidelity, or the raw weights?

GLM-5.2 NVFP4 keeps all 256 experts for the reference serve; BF16 gives you full precision to fine-tune evaluation and detector models.

GLM-5.2 NVFP4 (Full) $499 → GLM-5.2 BF16 $499 →