Uncensored Kimi K3.1 Pro edit keeps abilities intact
An open-weights edit of Kimi K3.1 Pro strips refusal behavior without altering its underlying capabilities. It answers prompts the base model declines, though less often than the project's closed-weight variants. — huggingface.co
Qwen3 42B uncensored build runs locally
An uncensored Qwen3 build arrives as a quantized package that runs with the most common local inference tools. — huggingface.co
Veyra 2.1.6 powers real-time video enhancement
Veyra 2.1.6 turns local video, capture feeds, and game streaming into a live, GPU-accelerated enhancement pipeline on Windows. — github.com
Claude Code permission file leaks secrets
Claude Code's "don't ask again" approval saves the full approved command to a project file, and 1,122 public copies of that file hold real passwords or keys. Anthropic plans to prototype basic secret scanning to catch the leak. — apiiro.com
Lyber AI ships uncensored Kimi K3.1 Pro
Lyber AI's uncensored Kimi K3.1 Pro build keeps the base model's extended thinking and tool calling while answering requests the original declines. — huggingface.co
DeepSeek Harness plugin unlocks free models
A plugin for the DeepSeek Harness client grants free, unlimited access to frontier models without login or API keys, and the repository reached the top of GitHub trending. — github.com
Laplace-1-Vector judges agent actions on-device
Juneo Lab's 4-billion-parameter Laplace-1-Vector scores candidate actions and predicts their consequences before an agent acts. The open-weights model runs on the organization's own hardware, keeping data on-premises. — huggingface.co
OpenThai fine-tune beats larger Thai OCR
A Thai fine-tune of NVIDIA's Nemotron document parser converts pages to Markdown with tables preserved as HTML, and it edges out larger OCR models on Thai government documents. — huggingface.co
Goldman Sachs restricts Claude Code to engineers
Goldman Sachs limits Claude Code to its engineering staff, taking the coding agent away from traders and bankers who had used it to monitor client flows and build market models. The bank is still a lead underwriter for Anthropic's IPO. — thestateofai.com
QAdapt neural pre-decoder cuts Willow errors
QAdapt's noise-adaptive neural pre-decoder learns local error patterns on quantum processors and passes remaining corrections to a conventional decoder. With frozen weights, it lowers the error rate on Google Willow records and unseen noise conditions. — huggingface.co