Uncensored Versions of LongCat, Jamba2, and Qwen Released with Local Deployment Tools
LLMFan46 · reddit · 2026-08-01
A developer has released uncensored versions of several popular large language models on Hugging Face, providing both Safetensors and GGUF formats and significantly reducing their refusal rates.
- LongCat-Flash-Lite: Refusal rate dropped from 100% to 9%. A specific llama.cpp fork was adapted to support local execution.
- Jamba2-Mini: A 52B parameter hybrid Mamba model with refusal rates cut from 97% down to 4%.
- Qwen3.5 Series: Features a custom 27B version built using J-Wash tools, and a 9B version whose misleading 96% refusal rate was reduced to 11%.
More from Models
- NVIDIA's Spatial-IQ Benchmark Exposes MLLM 3D Counting Flaws (17.7% vs 82.1% Human Accuracy) — NVIDIAAI · 2026-08-01
- NVIDIA's Spatial-IQ Benchmark Exposes Multimodal Models' Flaws in 3D Reasoning — NVIDIAAI · 2026-08-01
- Qwen Releases UI-Agent Technical Report: Unifying Cross-Platform GUI Interaction — _akhaliq · 2026-08-01
- New DeepSeek Flash Version Released: Improved Capabilities and Unlimited Access — bindureddy · 2026-08-01
- KOL Critiques: Benchmarks Named 'Frontier' Are Frontier No More — HanchungLee · 2026-08-01
- Claude 3 Opus Shows Collusion and Threats in Vending Benchmark — VraserX · 2026-08-01