New AI enzymes, 700+ math breakthroughs, AI for DNA mutations, room temp semiconductors: AI NEWS

W2WorldPlay 2Unverified

Open-source interactive real-time world generator with memory for long-term consistency. Built on Alibaba's Wan 2.2, with a 4-step turbo model; code released.

MSMSFlowAvailable in EU

Text-to-human-motion generator that supports joint constraints. Code and training scripts released, about 1.1 GB, Apache 2 license.

K6Kandinsky 6.0Kandinsky LabAvailable in EU

Open-source video generator with Lite (3.8B) and Pro (29B) variants plus a latent super-resolution model. Limited to 5-second clips and text/image-to-video; judged decent but behind Minimax.

CIChatGPT intelligent UIOpenAIAvailable in EU

ChatGPT can generate full interactive interfaces to answer questions, and GPT-6 can start answering while still thinking. Rolling out to paid plans then free.

ALALoDLMAmazonAvailable in EU

Looped diffusion language model combining parallel token prediction with recurrent refinement loops. Released in 1.7B and 8B variants under a non-commercial license, with higher throughput and lower compute.

ANAlphaProtein NovoGoogle DeepMindAvailable in EU

AI that designs new enzymes from scratch, both structure and sequence, validated with AlphaFold 3. Lab-tested enzymes produce piperidine and break down DEHP; not yet peer-reviewed.

B1Botanic 1Google DeepMindUnverified

Model trained on DNA of 320 plant species, used with Gemma 4 as a scientist agent to rank candidate mutations. Ranked the correct mutation first among about 2,500 candidates 90% of the time.

OMOpenAI math breakthroughsOpenAIAvailable in EU

OpenAI posted 719 manuscripts from an unreleased internal model across about 4,000 attempted problems, about a quarter Lean-verified. Highlights include a partial Riemann hypothesis result, Hilbert's 10th over rationals, a matrix multiplication exponent of 2.25, and Birch–Swinnerton-Dyer progress.

EGEmbedding Gemma 2GoogleAvailable in EU

Small open multimodal embedding model for text, image, audio and video, under a billion parameters, running on phones. Apache 2 license and available on Hugging Face.

MLMistral Large 4MistralAvailable in EU

1 trillion parameter MoE with 52B active, weights promised at end of month. Creator found its intelligence and price uncompetitive versus leading open models.

GEGenIAMetaUnverified

Turns a photo or video into 3D objects by building on SAM 3D and iteratively checking the match against the original image. Code released under a non-commercial license.

RTRoom temperature magnetic semiconductors discoveryVals AIUnverified

A team of Claude Opus 5.5 agents ran quantum simulations and found two candidate antiferromagnetic semiconductor materials for room-temperature memory. Results are simulation-only, not yet experimentally confirmed.

FLFlashDexRetargetUnverified

Converts human hand demonstrations into robot hand motions with a single policy trained across motions and look-ahead frames. Beats prior methods; code released, training needs about 110 GB GPU memory.

I3Iris 3BSperid LabsAvailable in EU

Open-source 3B image model that works in pixel space without a VAE. Supports text-to-image and image-to-image and can be fine-tuned for depth or upscaling; quality behind leading open models. Apache 2 license.

GTGPT-6.1 TerraOpenAIAvailable in EU

Teaches simulated human bodies to move across difficult terrain using muscle and skeleton simulation. Trained on 9 hours of motion data and extendable with fine-tuned policies; Apache 2 code released.

COCoDimReconUnverified

Turns photos of a real environment into an editable, physics-ready 3D scene with deformable object properties (cables, sheets, solids). Code is coming soon.

DIDistSceneUnverified

Generates an entire 3D scene from a single image, with environment and objects in a shared coordinate system, then refines each object. Best on benchmarks; model and data to be released.

NBNano Banana 2.1GoogleAvailable in EU

Minor upgrade to Nano Banana for complex instructions, text rendering, up to 14 reference images and 4K output. Creator found quality behind GPT Image 2.5; free to try in Google Flow.

CHClaude Haiku 5.5AnthropicUnverified

Anthropic's smallest, fastest model, a big improvement over prior Haiku and beating GPT-6 Luna on most benchmarks shown. About 21 cents per task, cheaper than Sonnet 5.5 but not frontier in intelligence; available on the free plan.