Whoa! Google Gemma-4-31B has been cracked!

· HarmBench score: 93.7% (149/159), basically capable of doing whatever you want.
· Abliterated technology: uses a technique called "abliteration" to remove safety barriers without damaging the model's intelligence.
But the most explosive part is its underlying Gemma-4 native capabilities:
· 31B Dense, performance surpassing expectations: third place on the Arena AI open-source leaderboard, achieving the same performance as models with 600B parameters using only one-thirtieth of the parameters.
· Multimodal versatile vision: directly understands images, videos, PDFs, OCR, handwriting, charts, and can even analyze videos.
· 256K ultra-long context: can process entire "The Three-Body Problem" trilogy or large codebases in one go without segmentation, understanding everything at once.
· Native intelligent agent capabilities: natively supports Function Calling and structured JSON output, capable of calling tools independently to complete tasks.
About local deployment:
· License: Apache 2.0, completely free for commercial use, modifiable at will.
· Ecosystem: directly supported by Ollama, LM Studio, vLLM, llama.cpp.
Hugging Face has already made it available for download:

A 31B model that can read, understand ultra-long texts, and act as an intelligent agent—now it’s been cracked open.