⚡AI Search⚡

What a crazy week in AI! 🚀 • Lance • LiTo • Flash GRPO • ReactiveGWM • L2P • Carbon • LongCat avatar 1.5 • MegaASR • HY-MT2 • Marlin 2B • Qwen 3.7 • PanoWorld • Stable Audio 3 • FashionChameleon & more! Watch the full recap: youtu.be/pC6KHflGye0

Flash-GRPO makes video diffusion alignment way more efficient. Instead of training through every step, it learns from just one step > Better quality under low compute budgets > Stable from 1.3B to 14B models > Free & open source shredded-pork.github.io/Flash-GRPO.git…

This could be huge. Microsoft's new SkillOpt trains skills without changing model weights. > Works across GPT + Qwen models > Transfers across models + harnesses > Code available microsoft.github.io/SkillOpt/

MiniCPM5-1B is a BEAST. Tiny open LLM built for local agents > Small enough to run on-device > Think / No Think modes > Beats similar sized models in knowledge, tool use, coding, reasoning > Free & open source, Apache-2.0 huggingface.co/openbmb/MiniCP…

This is so useful. Gepa-viz is a live visualizer for prompt optimization. It makes prompt evolution inspectable and trackable > Watch candidates grow as a graph > Inspect prompts, diffs, feedback > Open sourced, MIT license github.com/modaic-ai/gepa…

NVIDIA's PiD is a new pixel diffusion decoder for high-res image models. It skips decode-then-upscale stop, making sharper outputs faster. > Directly generates 4K images > Up to 5.9x faster than SeedVR2 > Free & open weights research.nvidia.com/labs/sil/proje…

SCOPE is a new FPS world model for controllable game video. It turns gamepad inputs into playable scenes. > Free & open source > 10-DoF controls > Trained on 69K clips across 7 FPS games > Zero-shot to unseen scenes z2tong.github.io/SCOPE/

Generative Recursive Reasoning: a new probabilistic framework for AI reasoning. This enables AI to explore possible solutions, not just think longer. > Beats Sudoku & ARC-AGI baselines > Samples many latent paths in parallel > Scales via depth + width ahn-ml.github.io/gram-website/

Aurora is a new agentic video editing framework. It turns vague edit requests into model-ready plans. > Scores 87.9 on AgentEdit-Bench > Handles replace, remove, style & insert edits > Uses tools for refs, masks & prompts yongshengyu.com/Aurora-Page/

Heres the correct link meigen-ai.github.io/LongCat-Video-…

LongCat-Video-Avatar 1.5 is a major upgrade for AI avatar generation. > Better lip-sync + mouth-shape accuracy > Stable identities across long videos & multi-person scenes > Smooth full-body motion, singing, animation, & hand interactions > Outperforms HeyGen, Kling Avatar 2.0, & OmniHuman-1.5 in consistency tests Free & open source longcat-video.github.io/LongCat-Video-…

@nunopereira_x Thanks for flagging. Heres the correct link meigen-ai.github.io/LongCat-Video-…

MEGA-ASR is a new SOTA speech recognition model built for the real world. Trained on 54 harsh acoustic scenarios: noise, reverb, clipping, & more Uses RL + semantic reconstruction to recover speech even from heavily corrupted audio Beats top open & closed ASR models by up to ~30% in difficult environments Free & open source! xzf-thu.github.io/Mega-ASR/

AI research agents are getting serious. AutoResearchClaw is an autonomous AI research system that actually learns from failed experiments. > Multi-agent debate for better hypotheses + result analysis > Self-healing execution that pivots/refines when experiments fail > Verifiable reporting to avoid fake numbers + hallucinated citations > Human-in-the-loop control with 7 intervention modes > Cross-run memory that turns past mistakes into future safeguards Beats AI Scientist v2 by 54.7% on ARC-Bench. arxiv.org/abs/2605.20025

We are making our discount permanent! 🎉 Enjoy building with DeepSeek-V4-Pro and bring your innovative ideas to life! 🚀