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.
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.
Secure checkout via Stripe. One‑time, with the commercial + research license. You'll land on a confirmation with a short access form.
Drop your Hugging Face username at checkout. We approve your access to the gated research repo — usually within the hour.
Pull the weights and the deploy kit. The serve script and image recipe stand it up in your own environment the same day.
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.
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.
REAP drops ~31% of the experts while holding quality — so the compact research build runs on half the GPUs of the full model.
Multi‑token‑prediction speculative decode wired for W4A16 NVFP4 on Blackwell — up to 2.1× the tok/s for high-volume evaluation runs.
Hold whole attack corpora, long transcripts, or entire codebases in a single pass — natively, no rope hacks on your end.
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.