AI finally sleeps, Gemini 4 Carbon, ETH cracked, energy-based models, Anthropic IPO: AI NEWS

MOMotionBricksNVIDIAAvailable in EU

Real-time animation tool generating smooth character and humanoid robot motion between target poses, up to 15,000 fps with 2 ms latency. Trained on 350k motion clips; code and two ~800MB models (human, Unitree G1) released under NVIDIA open model agreement.

IDIDOLUnverified

Turns a single image into a 3D object using Gaussian splatting, with adjustable quality. Only 0.66B parameters, runs on 8GB VRAM, about 19 s on a 4090. Code and models released.

SCSceneBloomUnverified

Generates complete 3D rooms with sensible furniture layouts from text, using zones and Flux for furniture images. Code, LoRA adapters and 3D asset library released.

UOUnlimited OCR 2BaiduAvailable in EU

Document reading model that processes multiple pages together to keep context across long documents. Beats several larger open models on parsing benchmarks; only 6.9GB.

UNUniActionUnverified

Single model that generates video from an image, text and action controls (force, camera, hand or robot pose). Built on Wan 2.2; code, training and evaluation scripts released.

AUAutoSwarmHKUDSAvailable in EU

Open-source harness layer that coordinates multiple agent harnesses such as Codex and Claude Code to work in parallel with shared context. Works on Windows, macOS and Linux.

P1Paradee 1.2Available in EU

Tiny TTS model distilled from Kokoro, 8.5MB compressed and CPU-capable, with four preset voices and no cloning. Apache 2 licensed.

HYHy-SimulMTTencentAvailable in EU

Real-time Chinese-English translation with voice cloning. 30B MoE with ~3B active, claims lower error than human interpreters; Apache 2, quantized versions available.

Dataset of over 2,600 simulation-ready 3D objects (~490GB) with physical properties for robotics training, released on Hugging Face under CC-BY.

OSOpenAI scientific discoveriesOpenAIAvailable in EU

OpenAI announced AI-produced results: disproving the exponential time hypothesis, a proof of the local smoothing conjecture in all dimensions, and a proof of QCD/Yang-Mills existence (not the mass gap). Lean verification was released.

LALuma AgentsLuma AILimited in EU

Sponsor segment: Luma's preset workflows (virtual try-on, Relight Studio) and agents that plan and generate assets across a project.

MAMatAnyone-EUnverified

Video matting that separates foreground from background including hair detail; real time on 4K on an H200, needs ~3GB memory. Code and ComfyUI node coming.

OPOpen-H3-Regenerate-2KNVIDIAAvailable in EU

Upscales low-res videos from MiniMax H3 and others (LTX 2) to 2K, up to 4K output and 15 seconds. The creator found it changes details such as faces. Code released.

STStreamVibeUnverified

Proactive video agent that watches a stream and decides when to offer help based on your goal. Code and models promised within 7 days.

MUMUNDOUnverified

World model that builds an explorable, physics-simulated environment from images, with object properties and physics engine. Code and weights due end of November.

GTGPT-6.1 TerraOpenAIAvailable in EU

Mid-tier GPT-6.1 model between Luna and Sol, strong at computer use, cheap with low factual error rate. Below Sol and Sonnet 5.5 on some benchmarks but much cheaper; available in paid plans and API.

G4Gemini 4 CarbonGoogleUnverified

Google's frontier model, ranked number one on Artificial Analysis, with top accuracy, lowest hallucination, and Vending Bench leadership. Currently limited to an initial cohort of cyber defenders.

NNNVIDIA Nemotron 3.5 TranscribeNVIDIAAvailable in EU

Speech recognition model with controllable transcript style, 40 languages, speaker identification and word timestamps, up to an hour of audio. About 34B parameters with 4.3B active.

G4Genie 4Google DeepMindUnverified

World model turning text or images into interactive worlds with real-time control and now audio. Can pair with Gemini 4 for reasoning; limited to trusted testers, coming to Ultra subscribers.

PHPhysAg3DUnverified

Turns a single image into an interactive 3D model with movable parts and physical behavior, using 3D reconstruction, an agent for part analysis and physics simulation. MIT licensed code released.

Training framework with wake and sleep phases where the model consolidates memory and generates its own practice problems. Near-perfect long-context memory up to 10M tokens; early prototype.

ATAudio8 TTS Realistic 0.6BAvailable in EU

0.6B open TTS with zero-shot voice cloning from a few seconds of audio, 23 languages, 2.6GB, Apache 2.

S6SenseNova 6.8 Code2VideoSenseTimeAvailable in EU

Generates animated videos by writing code that renders each frame, similar to Manim, allowing element-level editing. Based on Qwen 3.8 27B, about 55GB.

Proof-of-concept image generator that learns an energy function and iteratively lowers it instead of denoising like diffusion. Quality is poor; 2.44GB model with code and training script released.

OVOverSeerMetaUnverified

World model simulating a shared consistent world for multiple agents using fixed extra camera views. Demos in Minecraft, Doom and real life; model not released, only dataset.

33B dense open model that beats similarly sized models including Qwen 3.8 and Gemma 4 on agentic and knowledge-work benchmarks. 66.3GB; code and model released.

LFLFM3Liquid AIAvailable in EU

Family of six ~2B models (text, vision, audio, omni, decision) released on Hugging Face in many quantized variants including MLX. Text model outperforms larger competitors.

A3Apex 35BApexinUnverified

Open model based on Qwen 3.6, tuned for automated scientific research and writing robot-control code in simulation. Outperforms its base on research and agentic benchmarks; full model 72GB, smallest quant 9GB.

Family of four open MoE models from 19B (3B active) to 280B (16 active), pruned and fine-tuned from DeepSeek 4.1 and Qwen 3.6 35B. The small one beats Qwen 3.6 at half the size; FP4, FP8 and GGUF versions available.

Diffusion language model derived from Qwen 3.6 35B that generates blocks of tokens in parallel. Strong on agentic and tool-use tasks with favorable speed; full model 69GB.

SPSpark-X2.6-1.7BAvailable in EU

1.7B on-device model that beats similarly sized models such as Qwen 3.5 2B and Gemma 4 on many benchmarks and is fast on Mac M5 Pro and RTX 4090. About 3.4GB.

S9SenseNova-Loop 9BSenseTimeAvailable in EU

9B looped multimodal model that repeats layers for more computation without more parameters. Competitive with larger models on agentic benchmarks; Apache 2, about 38GB.