AIAny
AI Model2026
Icon for item

GLM 5.3 CRACK — Cybersecurity FP8

Provides a cybersecurity-focused CRACK variant of GLM-5.3 FP8 that reduces refusals for offensive-security, red-team, exploit-development and malware-analysis queries while retaining native FP8 speed on Hopper GPUs; MIT-licensed for authorized security work.

Introduction

Most general-purpose LLM safety edits trade off too much helpfulness for legitimate security researchers. This FP8 CRACK variant intentionally shifts that balance: it preserves GLM-5.3’s capabilities while reducing refusal behavior specifically for offensive-security and adjacent technical queries, so red-teamers and malware analysts can get more actionable, technical responses in authorized contexts.

What Sets It Apart
  • Targeted refusal reduction: tuned to increase direct compliance on cyber-offense probes (≈89% true-compliance on the cyber_offense subset of HB-320), while keeping hard-refuses near zero for the intended domain. This makes the model more likely to answer exploit-dev, evasion, phishing, reverse-engineering and credential-attack style prompts when used appropriately.
  • Native FP8 performance and architecture parity: based on an FP8 quant of GLM-5.3 with residual-writer edits (no LoRA/fine-tune runtime hooks), enabling tensor-core FP8 throughput on Hopper (H100/H200) and compatibility with vLLM serving patterns (example configs support TP8 on 8× H200 and up to ~131k context under recommended concurrency).
  • Measured capability preservation: logit-mode MMLU performance is effectively preserved (+1.07 pp vs base reported), indicating the capability surface remains close to upstream GLM-5.3 while behavior on safety probes is rebalanced.
Who It's For and Trade-offs

Great fit if you are an authorized red-team, penetration tester, CTF competitor, defender doing deep malware analysis, or an instructor building security labs and need fewer soft-refusals for technical cyber queries. Look elsewhere if you need a universal uncensor (there is a separate uncensored sibling) or if you require verbatim copyright reproduction—this variant still soft-refuses exact copyrighted text. Also avoid using it for unauthorized attacks, critical-infrastructure targeting, or any activity that would violate local law; the model authors explicitly state those limits and license it under MIT but place responsibility on users.

Where It Fits

This model occupies a narrow niche between stock GLM-5.3 FP8 and broad uncensored forks: it deliberately relaxes safety on cyber-technical prompts while aiming to keep overall model capabilities and FP8 performance intact. Use it when you need a defensible, reproducible model artifact that yields more permissive cyber-domain answers under controlled, authorized workflows.

Information

Categories

More Items

Hugging Face
AI Model2026

NVFP4-quantized checkpoint of Qwen3.8-Flash-Next for GPU-optimized multimodal autoregressive inference — routed MoE experts in W4A4 NVFP4 while attention/ancillary layers remain BF16; ~2.7× smaller than the BF16 source and supports very long contexts.

Hugging Face
AI Model2026

Fine-tuned variant of Qwen3.8-27B optimized to reduce reasoning cost and wall-clock latency for long-running agent workloads. Delivers ~12.8% faster decoding and higher MTP draft acceptance while cutting runaway reasoning, at a small MMLU-Pro accuracy trade-off (−1.45 pp).

Hugging Face
AI Model2026

Sparse MoE causal LLM that uses Mixture-of-Value Attention (MoVA) to store 36B parameters while activating ~4B per token; supports a native 524,288-token context and is released with final checkpoints, training data, and training code under open license.