Blackfrost AI · The Void Series · AI Security Research

GLM‑5.2

Red-team baseline · 0/450 512K context MTP · Blackwell NVFP4 · BF16

A refusal-ablated GLM‑5.2 built for AI security research — red-team your guardrails, benchmark robustness, and study model behavior at the boundary. Pruned and quantized on Blackwell iron so it runs on infrastructure you control, with the weights, deploy kit, and serving recipe included.

ALL SALES FINAL — NO REFUNDS. Digital goods delivered instantly via gated repository access. By purchasing you consent to immediate delivery and waive any right to cancel or refund. Refunds only where required by law or where we fail to deliver. See Terms.
Pick your build

Three ways to run GLM‑5.2.

The same refusal-ablated GLM‑5.2 research base in every build. The difference is footprint — how many cards it needs and how much precision you keep.

Best value · 4 cards
NVFP4 · REAP‑NU176
GLM‑5.2 REAP
31% of the experts REAP‑pruned, near‑lossless — the full research base on half the GPUs.
$399
one‑time · commercial + research license
  • Hardware4× RTX PRO 6000
  • Throughput~52 tok/s (MTP)
  • Context512K
  • Refusal bench0 / 450
  • Weights~302 GB NVFP4
Full · NVFP4
GLM‑5.2 NVFP4
All 256 experts, max fidelity at 4‑bit. The reference research serve.
$499
one‑time · commercial + research license
  • Hardware8× RTX PRO 6000
  • Throughput~56 tok/s (2.1×)
  • Context512K → 1M
  • Refusal bench0 / 450
  • Weights~420 GB NVFP4
Full precision
GLM‑5.2 BF16
Reference weights — for fine‑tuning your own evaluation and detector models.
$499
one‑time · commercial + research license
  • Precisionbfloat16
  • Best forfine‑tune / quant
  • Context512K → 1M
  • Refusal bench0 / 450
  • Weights~1.5 TB BF16
Guided access

Checkout to serving, in three steps.

Purchase

Secure checkout via Stripe. One‑time, with the commercial + research license. You'll land on a confirmation with a short access form.

Verify & access

Drop your Hugging Face username at checkout. We approve your access to the gated research repo — usually within the hour.

Deploy

Pull the weights and the deploy kit. The serve script and image recipe stand it up in your own environment the same day.

Why security teams use it

Measure the model, not its guardrails.

To red-team a safety system, train a content classifier, or benchmark how a model behaves under adversarial pressure, you need a baseline that doesn't refuse — so your results reflect the model itself. That's what these builds are: a controlled, refusal-ablated research base you run in-house.

A clean red-team baseline

Refusal-ablated so guardrails don't sit between your test and your measurement. 0 refusals across a 450-prompt benchmark — a reproducible control for evaluating defenses.

Near‑lossless REAP prune

REAP drops ~31% of the experts while holding quality — so the compact research build runs on half the GPUs of the full model.

MTP speed fix baked in

Multi‑token‑prediction speculative decode wired for W4A16 NVFP4 on Blackwell — up to 2.1× the tok/s for high-volume evaluation runs.

512K context

Hold whole attack corpora, long transcripts, or entire codebases in a single pass — natively, no rope hacks on your end.

What went into it
GPUs used
ClassRTX PRO 6000
Prune → quant → tuneweeks
Your cost to run4 cards
Powered by SatchelLM

The marketplace for research-grade, self-hostable models.

GLM‑5.2 is the first drop. Soon, SatchelLM opens to researchers and creators building self-hostable models for security and evaluation work — set your price and sell them here. We handle checkout, license, gated delivery, and the storefront.

Get GLM‑5.2 Creators — join the waitlist