Weight space programming, new 3D models, Gemini Omni 1.2, GPT Image 2.6: AI NEWS

HAHairDeckAvailable in EU

Generates complete 3D strand-based hair with joints from a single character image, preserving colors and animating naturally. Code and a ~2.45 GB model released under Apache 2, runs on Nvidia and Apple Silicon.

RERelayVSRAvailable in EU

Open video upscaler giving 4x resolution. A large model enhances key frames and a smaller model propagates detail to in-between frames. Based on Alibaba's Wan, ~10.2 GB.

N2NitroFusion 2Available in EU

Image model based on Z-Image that generates in one or two diffusion steps. Model and code released, 12 GB BF16 plus compressed versions, runs in ComfyUI or Diffusers.

STSTRATANvidiaAvailable in EU

Weather simulation model that predicts global atmosphere at ~5 km resolution in 10-minute steps with only 332M parameters. Released on Hugging Face (1.6 GB) under Apache 2.

FLFlowHMRUnverified

Reconstructs physically realistic 3D human body motion from regular video, usable for humanoid robot transfer and character animation. Code, models and a Hugging Face demo released.

REReLiftNvidiaUnverified

Tracks points in 3D from regular video, including through occlusions, by combining existing trackers and depth estimators with a correction model. Code, weights and training scripts released.

GOGemini Omni 1.2 FlashGoogleUnverified

Google's video model generating up to 15 s clips from text, image, video or audio input, with better reference consistency (virtual try-on), natural-language video editing and motion transfer. Price cut 30% to 7 cents/sec at 720p; available in AI Studio, API and Flow.

GIGPT Image 2.6OpenAIAvailable in EU

OpenAI's new image model adds animated GIFs, image markup, SVG export and is 50% faster. API has Flare (speed, ~6 s) and Sunburst (quality/consistency) variants; available in ChatGPT and Codex.

MIT research where GPT-6 Astra inspects and edits another model's weights without training, e.g. raising Qwen 3.5 0.8B's MATH-500 from 7.9% to 48%. Creator notes the gain mostly fixed repetitive reasoning loops, used tools, and failed to improve Gemma.

LLLucid LyraNvidiaUnverified

Turns a single image into an explorable 3D scene with recognized, editable individual objects, using a video model to predict novel views. Code is coming soon.

PIP-H3 ImagePruna AIAvailable in EU

Image generator/editor adapted from the Minimax H3 video model, 8 steps instead of 50, up to 16 MP, combines up to eight reference images and has a prompt enhancer. Weights on Hugging Face in BF16, 8-bit and 4-bit.

Open-source robotics toolkit with four libraries: RoboKit (motion planning), RetargetingKit (human-to-robot motion), GraspKit (grasping/policies) and RoboSkills (agent workflows). Code released on GitHub.

MOMOSS-LiveUnverified

Real-time multimodal voice assistant that handles screen/camera input, interruptions and speaker separation, with foreground conversation and background reasoning. Available via online demo and API; no weights announced.

PJPJiTLinum AIAvailable in EU

2B-parameter pixel-space image generator trained from scratch for about $21,000 of compute. Experimental; struggles with complex scenes and text. Released under Apache 2 (8.74 GB).

COCoLa3DUnverified

Turns a single image into a 3D scene with each object as its own mesh, processing the whole scene at once. Supports automatic, point and mask modes; code, training code and dataset released.

Tiny ~51M-parameter vision model (22 MB with 2-bit quantization) that detects objects, reads text and understands spatial relations, running in ~0.1 s on a phone. Model and code released.

DRDreamXUnverified

Video diffusion model that transfers motion from one character's video to another subject, including human hand to robot, and can translate to robot actions. Code coming soon.

ZAZAPUnverified

27B vision-language-action robot policy trained entirely in simulation with RL on 40 simple tasks, yet performs unseen tasks on real robots, matching GPT-6 Astra's success rate faster and cheaper. Only a paper released so far.

HYHy4TencentUnverified

Tencent's new model with visual understanding, computer use, 3D/Blender and knowledge-work abilities, claimed among the best open models. Weights due in coming days; available via Tencent Cloud and WorkBuddy.

753B-parameter open scientific model based on GLM 5.3, further trained on literature, lab records and full research trajectories including failures. Released in full, FP4 (~464 GB) and MX versions; sometimes beats GPT-6 Astra on science benchmarks.

GCGPT-6 CyberOpenAIUnverified

Cybersecurity-specialized model that found 81 unknown high-severity Firefox vulnerabilities and helped build exploits for 900+ Linux vulnerabilities. Access limited to approved researchers via the Daybreak program.

MUMuseMetaUnverified

Meta's personal AI assistant gains real-time voice conversations and expressive avatars, and can work on tasks while talking. Rolling out in the US on mobile, web and Mac.

MOMobiusUnverified

Single architecture for text-to-motion generation, motion captioning, completion and editing, outperforming competitors on benchmarks. Code is under internal review.

KKKimi K3.1Moonshot AIAvailable in EU

Best-performing open model; weights now released on Hugging Face. 2.8T total parameter MoE with 104B active, 1.56 TB in size, aimed at enterprises.

E5ERNIE 5.5 FlashBaiduAvailable in EU

Open 280B MoE base model with 14B active and 1M-token context, strong among similar-size open models. Weights in FP8 (~288 GB) and BF16 (~573 GB); a base model intended for further training.

Multimodal 560B MoE (25B active) accepting text, image and video, with spatial/3D understanding and agentic abilities. Currently available via Ant's developer platform; weights not yet released.

R6Ripple 65B PreviewRecursive IntelligenceAvailable in EU

Dense 65B open model trained from scratch for ~$32M, claimed state-of-the-art for agentic coding among US open models, with 1M context. Released on Hugging Face under Apache 2 (130 GB BF16, 57 GB FP4).

L2LongCat 2.5MeituanAvailable in EU

Multimodal 1.6T MoE (48B active) strong at coding, visual understanding and autonomous tasks. Available via LongCat Chat and API; open-source release marked coming soon.

I2Inkling 2 SmallThinking MachinesAvailable in EU

594B MoE (17B active) model strong at coding, visual understanding and long autonomous tasks, including autonomous AI research. Available on Tinker; weights planned next week under Apache 2.

VOVoxLutionUnverified

Generates fully furnished 3D indoor scenes from a category or simple layout, building coarse voxels then progressively refining detail. Code coming soon.

Reconstructs 3D geometry and camera poses from photos or video using a loop transformer, refining incrementally and supporting streaming video. Code and models promised within a week.