GPT 5.6, Mythos ban lifted, realtime avatars, Seedance 2.5, brain ultrasound: AI NEWS

WSWan StreamerAlibabaUnverified

Real-time conversational avatar handled by a single transformer interleaving text, audio and video. Runs at 25 fps with ~200 ms latency and duplex communication; v0.1 is 192p proof of concept, paper only, no code.

DSDomain ShuttleUnverified

Reference-to-video generator that keeps multiple characters and objects consistent, even across mixed art styles (realistic, 2D, 3D). Code released; built on a 1.2 base model, 35 GB, needs a high-end GPU.

O1Ornith 1.0Available in EU

Open-source MoE family (9B to 397B) for agentic coding that learns to generate its own harness/scaffold. Reported benchmarks are on par with larger open models like DeepSeek V4 and GLM; weights and GGUF quants are on Hugging Face.

ARArborStability AIUnverified

3D generation control system that takes constraint meshes specifying where geometry should or shouldn't exist and feeds them to existing generators like Trellis without retraining. Small 600 MB model, code released, usable as a Blender add-on.

OpenAI's first inference chip, made with Broadcom, developed in 9 months with help from OpenAI models. Early testing shows better performance per watt than current accelerators; gigawatt-scale deployment planned.

ISIBM sub-1nm chip technology (nanostack)IBMUnverified

0.7 nm (7 angstrom) node using a 3D nanostack architecture that vertically stacks transistors, fitting nearly 100 billion transistors on a fingernail-sized chip. Projected up to 50% more performance or 70% better efficiency versus the 2 nm node.

DADanceOPDByteDanceUnverified

Training framework that distills multiple specialist teachers (one per training example) into a single model for text-to-image, editing and style transfer. Scores highest on combined benchmarks; paper out, code coming soon.

S2Seedance 2.0ByteDanceUnverified

Higgsfield (sponsor) added 4K generation for Seedance 2.0, with multi-shot scenes, character consistency and multimodal references. This is a sponsored segment.

Image generator using coupled oscillators instead of diffusion. Image quality is poor and it's a proof of concept; weights and training scripts are open-source under MIT.

H5HIW 500BitRobot FoundationAvailable in EU

Open dataset of 500+ hours (10+ TB) of whole-body teleoperated Unitree humanoid data in real homes across 10+ household tasks, with head and wrist cameras. Raw data released; LeRobot format coming.

S2Seedance 2.5ByteDanceUnverified

Unveiled upgrade to Seedance 2.0 with 30-second clips, up to 50 multimodal references, local box-based editing, stronger consistency, native audio and 2K/4K output. Launch expected in early July.

S2Seed 2.1ByteDanceUnverified

Closed-source, paid multimodal model family (Pro and Turbo) positioned as a productivity agent. Roughly on par with Claude Opus 4.7 and GPT-5.5 on work and coding benchmarks, and stronger on chart and video understanding; free in their studio and via API.

LILift4DUnverified

Turns a single ordinary video into a full 4D scene (3D plus time), reconstructing regions the camera never captured. Code is coming soon.

PEPerceptionDLMByteDanceAvailable in EU

Vision-language model using a diffusion language model to caption multiple image regions simultaneously. Outperforms similar-sized vision models and is faster; models, data and code are open source.

HHHappy Horse 1.1AlibabaUnverified

Video model upgrade with better motion realism, consistency and native audio/lip sync. Supports 1080p, 3-15s clips, text/image/reference-to-video (up to 9 references), wide aspect ratios and multiple languages; still behind Seedance.

K2Krea 2KreaAvailable in EU

New top open-source image generator with realistic images, strong prompt adherence, world knowledge and very little censorship. Light enough for about 4 GB VRAM and about 10 seconds per image for the creator; a full tutorial was previously made.

DRDeepSeek R1DeepSeekAvailable in EU

Unitree's lightest and cheapest humanoid, starting at $4,900, shown in a demo doing acrobatic breakdancing moves, flips and stable landings.

GSGPT-5.6 SolOpenAILimited in EU

Model family (Sol, Terra, Luna) with a new max reasoning effort and an ultra multi-agent mode, setting a new Terminal Bench state of the art and beating Mythos there, though weaker than Mythos on cybersecurity. Up to 750 tokens/s on Cerebras; only a limited preview for trusted partners under US government compliance.

After the US government banned access to Claude Fable and Mythos, a letter eased the ban to allow Mythos 5 release to about 100 trusted partners including US companies and agencies. The creator argues this shows the importance of open-source models.

Ultrasound method with injected microbubbles that maps blood flow in the living human brain in 3D, claiming 100x the resolution of comparable CT scans. The pipeline and datasets are open-sourced; the creator notes the microbubble injection is not fully non-invasive.

AUAutodataMetaUnverified

Agentic framework that autonomously builds and iteratively improves training and evaluation datasets using a challenger agent and rubric in a loop. A Qwen 3.5 4B model trained on its data beat a standard method; a technical report is released.

SFSakana FuguSakana AINot in EU

Paid, closed orchestrator that routes prompts across closed and open models, claimed to beat frontier models and match Fable 5 and Mythos preview. The creator calls the claim misleading since it's an ensemble router rather than a base model, and criticizes the pricing and lack of transparency.