Technical AI news, automatically curated and generated
A 9B open-weight model was fine-tuned using Reinforcement Learning (RL) at a cost of $500 to specialize in catalog reviews. This specialized model outperformed frontier models on this specific domain task. The experiment…
→ View original source
Moonshot AI has released Kimi K3, a 3-trillion-parameter open-source mixture-of-experts (MoE) model featuring a 1M-token context window. This development highlights a shift in the AI industry toward open weights, aggress…
→ View original source
A pull request by wjinxu adds DSpark speculative decoding support to llama.cpp (PR #25173). DSpark leverages multiple draft models for speculative decoding, with reference implementations including DeepSeek-V4-Pro-DSpark…
→ View original source
Modern site reliability requires handling massive cross‑service context, and effective AI integration
→ View original source
Google Cloud has introduced the Gemini Distillation Service, a feature designed to streamline the model optimization process. This service allows users to leverage larger models to create smaller, more efficient versions…
→ View original source
Moonshot AI’s Kimi K3, a 2.8 trillion‑parameter model, was released with fully open weights and an
→ View original sourceReasoning-Medical-27B is a fine-tuned Qwen3.6-27B model optimized for advanced medical reasoning, trained on 370,000 high-quality Q&A examples with Chain-of-Thought reasoning. It supports universal medical knowledge, inc…
→ View original source
Claude Opus 5’s prompt leak, containing 34,000 tokens, was inadvertently published and has been hailed as the premier prompt‑engineering reference of 2026. The accidental release offers unprecedented insights into high‑p…
→ View original sourceKimi AI and kvcache-ai have open-sourced AgentENV under the MIT license to enhance environment throughput for agentic reinforcement learning (RL) training. Designed for Kimi K3, this distributed system utilizes Firecrack…
→ View original source
A Hacker News post warns that shared Claude chats and Artifacts may have been crawled and indexed by Google, potentially exposing user data. The issue raises concerns about privacy and data handling practices of the AI s…
→ View original source
Jensen Huang said that during the Hugging Face incident, closed AI blocked essential forensics. An open-weight frontier model helped contain the intrusion, which led to the creation of the Open Secure AI Alliance. Read o…
→ View original source
On-policy diffusion distillation (OPD) extends velocity matching to classifier-free guidance (CFG)-composed predictions, but its branch-level behavior remains under-identified. The study reveals that directly matching te…
→ View original source