New Deepseek, human genome map, Navier Stokes, GPT finance, Suno v6, YuE2: AI NEWS

Pixel-level model that converts a regular image into depth maps, surface normals and albedo. Beats Moge 3 and InfiniDepth in comparisons and is released; needs about 17 GB VRAM at one resolution and 29 GB at higher.

UNUnimateUnverified

Animates rigged 3D models with arbitrary skeleton types (non-human characters) from a text instruction, without retraining. Code and training code are open source.

Database of AlphaGenome predictions for every possible single-letter human DNA mutation (over 9 billion variants, ~1 PB). Includes an impact score and found 22% more genetic associations in UK Biobank data. Publicly searchable.

LWLingbot World 2RobbyantUnverified

Real-time interactive world model controlled by key presses and prompts, running at 720p 60fps and generating worlds for over an hour. Adds NPC agents, built on Wan 2.2, with 14B and 1.3B variants released.

I0Isaac 0.5PerceptronUnverified

Open-source robot foundation model taking images, video, language and robot state to predict actions or answer questions. Trained on 35+ robot systems, 100k hours of robot data and about a million hours of video.

WSWorld SculptUnverified

Turns multiple images or a video into a 3D scene with separate editable objects, and can convert Gaussian splat worlds into mesh scenes. Code released.

FIFire3DUnverified

Reconstructs simulation-ready 3D scenes from photos or video, with each object as its own mesh with materials. Runs in under a minute and handles 16 objects in parallel on one GPU; code and training code released.

AUAuKTencentAvailable in EU

Speech model described as Nano Banana for speech. Does text-to-speech from voice descriptions, zero-shot voice cloning, word-level edits, denoising, enhancement, emotion and timbre changes, whisper conversion. Model is about 6.12 GB and released open source.

DVDeepSeek V4.1 FlashDeepSeekAvailable in EU

552B-parameter MoE open model with 8-16B active, new architecture with encoder-decoder, Engram and sliding window attention. Top open model on several benchmarks, 217 tokens/s via API, very cheap; weights about 510 GB with GGUF quantizations available.

ONOpenAI Navier-Stokes solutionOpenAIUnverified

OpenAI's internal model, using 10,000 concurrent agents over 88 hours, found a finite-time blowup solution for the forced non-Euler Navier-Stokes equations. The unforced problem remains unsolved.

SHShow HarnessUnverified

Harness that lets any vision-language model (Gemini, GPT, Qwen, GLM) control robots by choosing from named actions, converted by an action interpreter. Success reached up to 100%; requires a real robot.

EDEdge0Available in EU

Framework that streams only needed experts of MoE models into memory, running Qwen3.5-35B in about 2.9 GB and an 8B variant in 1 GB, with a Recover LoRA to regain int4 quality loss. Runs on Mac.

RERealSWEAvailable in EU

Benchmark giving agents access to private company code for business-relevant changes. Top scores are under 40% (Fable 5.1 at 38.8%, GPT6 Astra 33.8%), with the top four statistically tied.

SVSuno V6SunoAvailable in EU

Latest Suno music model with finer control and micro-editing such as changing one lyric or combining vocals and instruments. Comes as V6 and V6 Wild for paid plans and V6 mini for free users. Closed and paid.

YUYuE2Multimodal Art Projection (M-A-P)Available in EU

Open-source music generator that first composes an editable musical plan (melody, chords, structure) then renders a full song with vocals. Claims to beat MiniMax Music and ACE-Step and even Suno V6; model is 7.3 GB and supports covers.

CFChatGPT for Financial ServicesOpenAIUnverified

Combines GPT6 Astra with financial data (PitchBook, Crunchbase) to research companies, build models and produce spreadsheets and presentations with traceable figures. Access requires contacting sales.

UMUMRUnverified

Retargets human motion to humanoid robots by representing body surfaces as 3D point sets and learning correspondences, preserving contacts. Handles spin kicks, stairs and tennis; code released.

UNUnifoLM-WMAUnitreeUnverified

Open-source 6B-parameter robot model covering 64 tasks including whole-body coordination, supporting different grippers and hands. Predicts scene changes and generates actions; models and dataset released, under 9 GB.

M2MiniCPM5 2BOpenBMBAvailable in EU

2B dense small model that is state-of-the-art for its size, beating Qwen3.5 4B on average across benchmarks. About 5 GB; the training dataset is also released.