AI scientist, DNA editors, AI NPCs, new Qwen, open-source robots, new video editors: AI NEWS

LALanceByteDanceUnverified

3B-parameter unified multimodal model that does text-to-video, video editing (multi-turn), text-to-image, image editing and visual understanding. Code released; needs a GPU with at least 40 GB VRAM.

LILiToAppleUnverified

Generates 3D models from an input image while capturing view-dependent appearance like reflections and lighting. Reported as more accurate than Trellis; code and training scripts released.

FGFlash GRPOUnverified

Method to align video models with human preferences by sampling a single timestep, using isotemporal grouping and temporal gradient rectification. Learns faster than Flow GRPO fast and improves detail and motion; code released.

REReactiveGWMUnverified

Reactive game world model where NPC behavior can be steered by high-level strategy prompts separate from player actions, injected via cross attention. Built on Wan 2.2 and code is released.

L2L2PAvailable in EU

Pixel-space diffusion image model built from Z-Image by removing the VAE and latent space. Handles up to 4K and 8K extrapolation; only the 1K model (~20 GB) is released so far, with training scripts.

Open-source DNA foundation model processing nearly 400,000 base pairs, able to continue sequences, score variants and predict protein structure. Claimed 275x faster than Evo 2 medium; sizes from 500M to 8B parameters.

LVLongCat Video Avatar 1.5MeituanAvailable in EU

Talking avatar generator from a reference image plus audio, more stable and expressive than v1, supports animation styles and multi-person conversation. Weights released; int8 version is 16 GB.

MAMega ASRUnverified

Speech recognition model for noisy real-world audio, trained on 2.6M samples across seven acoustic problems. Claims nearly 30% gains over leading models in difficult conditions; under 5 GB with fine-tuning scripts.

HYHY-MT2TencentAvailable in EU

Family of open multilingual translation models (1.8B, 7B, 30B MoE with 3B active) supporting 33 languages and detailed instruction following for formatting. Outperforms larger open models on benchmarks; FP8 and GGUF versions released.

ACAI co-scientistGoogle DeepMindUnverified

Multi-agent system that generates hypotheses, reviews literature and proposes experiments through internal debate among agents. Published in Nature, with examples in drug discovery such as liver fibrosis.

2B video language model, based on Qwen 3.5 2B, that produces scene descriptions and timestamped events. Reported as matching Gemini 2.5 Flash on captioning; under 6 GB.

Q3Qwen 3.7 MaxAlibabaUnverified

Latest Qwen aimed at agentic and coding tasks with vision capabilities, on par with top open models like DeepSeek V4, GLM 5.1 and Kimi K2.6 on benchmarks. Available via Alibaba Cloud Model Studio and API, not open-sourced yet.

Q3Qwen 3.5 LiveTranslateAlibabaUnverified

Live speech translation model that uses visual context to improve translation, supporting 60 languages with early streaming translation. A free online demo is available.

RPRobot Plus Plus magnetic wall-climbing robotRobot Plus PlusUnverified

Teleoperated magnetic wall-climbing industrial robot with humanoid dual arms for welding, grinding, inspection and painting on tanks and ship hulls. Claimed to have serviced over 10,000 ships.

Open-source 3D-printed humanoid robot platform with design, parts list, assembly guide, simulation and training software. Costs around $2,500 in parts.

UGUnitree G1 voice controlUnitree RoboticsUnverified

Demo of the G1 humanoid following spoken commands in real time in one continuous shot, performing jumps, planks, squats and a dance autonomously.

COCogOmniControlUnverified

ControlNet-like system for video generation using sketches, pose skeletons or line art plus a reference image and text prompt. Only a paper is released so far.

Video-to-audio model that generates sound directly in raw waveform space without a VAE or latent compression. Competitive with MMAudio, but the full production checkpoints were not released; piano example was poor.

PAPanoWorldUnverified

Generative world model that turns a floor plan and style reference into consistent panoramic tours of a full house, using a 3D shell as memory. Code and models not yet released.

SAStable Audio 3Stability AIAvailable in EU

Family of audio generation models with open-weight small and medium (1.4B, up to 6:20) variants, a small SFX version, and an API-only large model. Supports LoRA training and audio inpainting.

FAFashionChameleonAlibabaUnverified

Real-time video virtual try-on that switches garments during generation, running near 24 fps on a single GPU and claimed 32-180x faster than baselines. Code and model are planned but not yet released.