GLM-5.2 · NVFP4 · Full (256 experts)

GLM‑5.2 NVFP4

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.

Red-team baseline · 0/450 All 256 experts 512K → 1M MTP · ~56 tok/s
One-time · commercial + research
$499
Buy GLM‑5.2 NVFP4Bitcoin — $449 · save 10%
Secure checkout via Stripe
Gated research access approved after purchase
Overview

Maximum fidelity for the reference serve.

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.

Specifications
BaseGLM-5.2 (refusal-ablated)
ArchitectureMoE · DSA sparse attn
Experts256 / 256 (full)
QuantizationNVFP4 · W4A16 · group-16
Kept at BF16attention · lm_head · norms
Context512K (→ 1M, YaRN)
Throughput~56 tok/s · MTP 2.1×
Refusal benchmark0 / 450 adversarial
Weights~420 GB
Hardware8× RTX PRO 6000 (Blackwell)
Licensecommercial + research
What's included
  • Full NVFP4 weights — gated Hugging Face research download
  • Deploy kit: serve image recipe, launch scripts, and the W4A16 MTP speed patch
  • Think / no-think chat templates
  • Serving guide tuned for 8× Blackwell
  • Commercial + research license
Quickstart

Serve it with the included kit.

# from the deploy kit — 8× RTX PRO 6000
MTP=1 MAX_MODEL_LEN=524288 bash serve_glm52_nvfp4.sh

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

Max-fidelity red-teams

Eight-card serves where you want the full model with nothing pruned for your evaluations.

Safety-system evaluators

Benchmark guardrails and classifiers against an unclipped reference baseline.

1M-context research

Long-transcript analysis, whole-repo agents, and extended adversarial sessions.

Other builds

Only have four cards, or want the raw weights?

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.

GLM-5.2 REAP (4-card) $399 → GLM-5.2 BF16 $499 →