DeepSeek-V4 Flash With Vision Support Is Now Live
DeepSeek-V4 Flash, now live with vision support, marks a new multimodal milestone for the DeepSeek model family.
LTX Ripple: First-Frame Video Editing with an Efficient IC LoRA
LTX Ripple is a new efficient IC LoRA approach for video editing, built on LTX 2.5. Users edit only the first frame and the changes ripple through the rest of t…
MiniMax H3 Max Used to Build Real-Time Interactive AI Games
A user shows how MiniMax H3 Max can power interactive games where all decisions are up to the player, with the model's speed making delays nearly nonexistent.
Funding Call: Someone Wants to Back LLM Humor Projects
A brief but notable signal: a funder is looking for someone building funny LLMs, pointing to humor as an emerging AI application area.
Sliding-Window Attention with Attention Sinks Outperforms Linear Attention Post-Training
A new research result shows that switching to a sliding-window attention mask with attention sinks—at no additional cost—beats linear attention during post-trai…
RL-Trained Coding Model Paints with JavaScript; Open Artifacts Coming Soon
A developer shares that they used TRL and OpenEnv to train a coding model to paint with JavaScript, and plans to release a blog with code, models, dataset, and…
Tencent Hunyuan Hy4 Preview Creates “Qingming Riverside” Animation, Outperforming GLM5.3-Flash on Same Prompt
A user reports that Tencent's Hy4 preview model generated an animated version of the classic Chinese painting 'Along the River During the Qingming Festival' usi…
Hy4 Preview Impresses in One-Shot Test: “Don’t Sleep on It”
A tester ran a one-shot comparison with the same prompt and found Tencent's Hy4 Preview surprisingly competitive, calling it a huge leap over Hy3.
MiniMax H3 Max Powers Interactive Games with Near-Zero Delay
A user reports using MiniMax H3 Max to create interactive games where all decisions are user-controlled, with the model's speed ensuring no noticeable delays.
Command Code AI Claims 30% Token Savings with New Shell Tool and TEF Benchmark
Command Code AI is highlighting its shell tool as the most token-efficient harness, claiming a 30% reduction in token bills and a spot on the TEF Pareto frontie…