Self-improving AI, Opus 4.8, Nvidia bangers, game-ready 3D models, juggling robots: AI NEWS

LOLocateAnythingNvidiaAvailable in EU

Vision-language grounding model that detects and segments objects in images/video using parallel box decoding, predicting whole bounding boxes in one step. Trained on 103M queries and 785M boxes; 3B params, 7.8 GB, code and training released.

COControlLightUnverified

Image editing tool for brightening dark images with exact control over light level, using generative brightening that preserves detail. Built on Flux 2 Klein; code, training code and dataset released.

TRTriSplatUnverified

3D reconstruction model that turns images into simulation-ready scenes by representing them as triangle primitives instead of Gaussian splats, skipping mesh conversion. Models are about 4.4 GB and code is released.

PIPiDNvidiaAvailable in EU

Pixel diffusion decoder that upscales images to 2K+ with strong detail, replacing the latent decode plus separate upscaler step. Upscales 512 to 2K in under a second, about 6x faster than SeedVR2; works with Flux 2, Z-Image, SD3.

INInstructAV2AVUnverified

System that edits video and audio together from a prompt, changing what a person says along with their lip sync, or changing voice and gender. Code and weights are listed but the GitHub page is still empty.

GEGenReconUnverified

Turns smartphone videos or image sets of a room into a PBR-ready 3D mesh scene using chunked generation and Trellis 2 as a shape prior. Supports relighting and editing; only a paper is out, code marked coming soon.

SCScopeTencentAvailable in EU

Playable first-person-shooter world model that takes a starting frame and controller actions and generates real-time responsive video. Trained on about 70,000 clips from seven FPS games; beats Matrix Game 3 and HunyuanWorld; about 30 GB, code and dataset released.

POPhysX OmniUnverified

3D generator producing simulation-ready assets with geometry, scale, materials and articulated joints, unifying rigid and articulated objects in one framework. Outperforms Articulate Anything and PhysX-Anything; fully open source with training script and dataset.

DEDeepSWEDatacurveAvailable in EU

Benchmark for coding agents with tasks written from scratch across 91 open-source repos, short prompts and large reference solutions, checked with handwritten behavioral verifiers. GPT 5.5 scores highest, then Claude models, then Gemini 3.5 Flash, while open models like Kimi K2.6 and GLM score lower.

COClaude Opus 4.8AnthropicAvailable in EU

Anthropic's latest model, beating Opus 4.7 and GPT 5.5 on most self-reported benchmarks, with improved honesty. Independent leaderboards are mixed: first on Artificial Analysis by one point and slightly cheaper than GPT 5.5, but not top on omniscience accuracy or LiveBench.

Home humanoid robot with a wheeled base shown doing kitchen, laundry, ironing, bartending and industrial tasks. Rumored price is around $13,000, but it can't walk or climb stairs.

AZAthena ZeroRAI InstituteUnverified

Humanoid robot demo that learned to juggle and switch between five juggling styles after under 10 minutes of real-world interaction.

S3Step 3.7 FlashStepFunAvailable in EU

Open-source multimodal flash model for real-world agents handling images, interfaces, web search and long runs. Approaches GPT 5.5 and Opus 4.7 on SWE-Bench Pro; about 400 GB, needing a DGX Spark or multiple GPUs.

CUCubePartUnverified

Text-to-3D generator that decomposes objects into a user-chosen number of separate part meshes using a diffusion transformer with cross-part attention. Gives cleaner boundaries than other part-based generators; under 10 GB and code released.

RHRelightable Hollywood-ported CharactersGoogleUnverified

Method that captures a moving human with multiple cameras and builds a full-body avatar that can be placed in new scenes with realistic relighting. Code is not yet released.

BBBES (Bidirectional Evolutionary Search)Available in EU

Technique combining forward evolutionary search that recombines partial attempts with backward goal decomposition into subgoals to give denser feedback. Shows gains on hard post-training tasks; the creator says the 'self-improving' label is a bit misleading. Code released.

AUAutoScientistsUnverified

Open-source multi-agent framework where agents act as a decentralized research lab with a shared state, experiment log, forum and dead-end registry. Beats other agentic frameworks on BioML-bench's 24 biomedical ML tasks.

GWGamma WorldNvidiaUnverified

Multi-agent world simulator that generates real-time 24 fps video of two to four players in a shared environment, using simplex rotary agent encoding. Code, training scripts and dataset tool are coming soon.

P3Pantheon 360Unverified

360-degree video generation model for digital twins that takes several panoramas plus a camera path and uses a reconstructed 3D point cloud to keep the panoramic video consistent. Code is coming soon.

1-bit and ternary compressed versions of Flux 2 Klein, shrunk from about 8 GB to about 1 GB to run offline on a phone. Generates a 512x512 image in 9.4 seconds on an iPhone 17 Pro Max.

M1MiniCPM5 1BOpenBMBAvailable in EU

Tiny 1B-parameter dense model, about 2 GB, that outperforms similarly sized models on knowledge, coding, math, reasoning and agentic benchmarks. Runs on laptops or phones.

SESegaUnverified

Method for generating very high-resolution images (4K to 6144 px) with sharp detail, working with Flux and Qwen Image bases. More consistent than other upscalers such as DyPE; code released for Flux 1 and Qwen Image.

PIPixlRelightUnverified

Single-photo relighting system where the user drags a light source or sets a spotlight. It estimates rough 3D scene geometry, relights it in Blender, and uses that as reference for the final prediction; code released.