LinkCantina SecurityTool
Security teams require unconstrained models to inspect vulnerabilities effectively, but must weigh altered refusal mechanics against unverified performance and reliability.
Cantina Security, in partnership with Yeta (@yetalabs), released cantina-security/apex-flash-1-abliterated as an experimental derivative of its open-weights model apex-flash-1. The checkpoint uses the GLM-5.3-Flash architecture with 321 billion parameters in BF16 Safetensors format, retaining the original image-text-to-text multimodal pipeline.
The derivative broadly modifies refusal behaviour across prompts rather than limiting adjustments to security-specific tasks. Cantina Security indicates that the checkpoint is intended solely for authorised security research in environments that researchers own or have formal permission to test. Crucially, the authors note that altering refusal mechanisms does not establish improved reliability or capability on security workloads.
No separate full-suite benchmark evaluation was run for this specific release. Consequently, performance metrics reported for the standard apex-flash-1 model cannot be attributed to this derivative, and visual capabilities remain unverified. The model can be served locally or in production via Hugging Face Transformers, vLLM, SGLang, and Docker Model Runner.
Summary generated by AI (Gemini 3.8 Flash, 9 November 2026). It may contain errors – the original source is authoritative.
Similar in content, found by the AI search.
X post:Cantina
Cantina has introduced apex-flash-1, an open-weights model designed for security research. It was post-trained on real vulnerabilities identified by the team, alongside an abliterated version designed to reduce refusals in authorized workflows.
1.5kbookmarks270.4kviews
Link:OrcaRouter
OrcaRouter has released an abliterated, block-FP8 quantized version of the multimodal model Qwen3.8-27B on Hugging Face. By orthogonalising the internal refusal direction, standard safety guardrails were removed specifically for red-teaming, interpretability, and AI safety research.
Link:Z.ai
Z.ai has launched GLM-5.3-Flash under the MIT licence on Hugging Face, marking the first natively multimodal model in the GLM-5 family. Featuring 320B total parameters and 18B active parameters, it targets high efficiency across coding and agentic tasks.
Link:PocketAiHub
PocketAiHub has released an unofficial MLX derivative family of Qwen3.8-27B with suppressed refusal behaviour ('abliteration'). The consolidated repository provides quantized builds ranging from experimental 2-bit AWQ to full BF16 optimised for Apple Silicon.
X post:O
In an online discussion on GLM-5.3-Flash, user O highlights community post-trained derivatives hosted on Hugging Face. UkisAI's Swift 1.5 significantly cuts reasoning tokens for general daily use, while Cantina Security's apex-flash-1-abliterated targets cyber research. The author ranks both variants ahead of stock GLM-5.3-Flash and DeepSeek.
13bookmarks468views
Link:Empero AI
Empero AI has released Qwythos-27B-v1, an open-weight reasoning model based on Qwen3.5-27B. The checkpoint preserves multi-token prediction (MTP), vision capabilities, and a 1M-token context window while specialising in terminal and agentic workflows.
Parts of this content were created with the assistance of AI.
