GPT 6 Astra, Claude Fable 5.1, Gemini 3.8, realtime Minimax, new world models: AI NEWS

HWH3 WorldUnverified

Turns the open-source Minimax H3 video generator into an interactive video game engine by translating key presses into short text descriptions fed to the model's language layer. Code and training script are released.

SOSolarWMUnverified

Framework converting video models into real-time interactive worlds that stay consistent for over an hour. Works with several video models, with a 1.3M clip, 25TB dataset, models and training recipes all open-sourced.

T3TimesFM 3Google ResearchAvailable in EU

Open-source, zero-shot, multivariate time series foundation model with 330M parameters, pre-trained on over a trillion time points. Ranked first among time series foundation models on several benchmarks.

LULucidaByteDanceUnverified

Turns images of a messy room into an editable 3D simulation by identifying objects, generating complete meshes for each and placing them back. Only a technical paper is released so far.

VIVideoDeltaNetLimited in EU

Replaces most of Minimax H3's attention with hybrid linear attention plus distillation for big speedups with minimal quality loss; can be faster than real time on eight B200s. Model, code and ComfyUI nodes are released.

LLLLaDA-ImageinclusionAIAvailable in EU

Open-source 6B text-to-image generator and editor producing realistic photos, posters and infographics. Comes in base (50 steps) and turbo (4 steps) versions in BF16 and FP8.

DVDeepSeek V4 Flash Vision ExperimentalDeepSeekAvailable in EU

Vision-enabled version of DeepSeek V4 Flash, said to be roughly on par with Claude Opus 4.8 in many cases. Now open-sourced on Hugging Face at 305B parameters and about 168 GB.

Q3Qwen 3.8 Max 0902AlibabaUnverified

Updated Qwen 3.8 Max with the same 2.4T-parameter architecture and 1M context, post-trained further on coding and co-work. Beats the previous version by large margins on agentic benchmarks; available via Qwen Cloud API, not open source.

Anthropic's newest model, strong at agentic scientific research, coding and knowledge work and top of an independent leaderboard. The creator flags very high cost, rapid usage-limit exhaustion and fallback to Opus 5 on research prompts.

G3Gemini 3.8 FlashGoogleUnverified

Fast Flash model that beats larger models on finance, legal and agentic terminal coding, and ranks first on Deep SWE. Uses many output tokens and ranks lower on LiveBench; a Cyber variant targets defensive security and vulnerability patching.

Meta's near-frontier model focused on long, messy agentic jobs and multitasking within one conversation. Available in Muse Code and the Meta model API.

GAGPT-6 AstraOpenAIUnverified

OpenAI's new model that uses the computer directly and posts state-of-the-art results on computer use, terminal coding, ARC-AGI-3 and FrontierMath. Rolled out to paid plans; the creator notes it ranks 2nd and 3rd on some independent indexes.

W3WeatherNext 3Google DeepMindUnverified

AI weather forecasting system using live satellite data to produce hourly global forecasts at up to 5 km resolution, with up to 60% better satellite precipitation scores. Being integrated into Search, Gemini, Maps, API and Earth Engine.

Complete brain and nervous system wiring map of a male fruit fly with over 166,000 neurons and about 125 million synapses, built from electron microscope slices and AI reconstruction.

ATAtlasWorld LabsUnverified

World model taking text, images, video or 3D input to create or reconstruct 3D worlds, with up to one minute of coherent 1440p video along user-drawn camera paths and outputs like point clouds and Gaussian splats. Currently a preview with early access form.

ILIntern Lumina U2InternLMUnverified

Open-source diffusion large language model that generates and edits images and also has strong image understanding, such as identifying scans, paintings and ER diagrams. Code instructions are out but weights are not yet on Hugging Face.

VAViggle AnimateViggleAvailable in EU

Swaps the character in a clip using one edited first frame, without pose estimation or segmentation. It is a fine-tune of Minimax H3 with released full weights (66 GB) and a 2.5 GB LoRA.

GWGWM Worlds 2RunwayUnverified

Real-time interactive world model producing 720p 24fps video and audio with no fixed maximum duration, driven by commands like walking or changing weather. Auto-regressive diffusion; research preview only, via a contact form.