New Deepseek, Seedance 2.5, Minimax H3, Gemini Robotics, AMD models: AI NEWS

IDID-V2VNetflix (Eyeline Labs)Unverified

Open-source video editing AI that changes the style of a scene (background, lighting, clothing) from one edited keyframe while preserving character identity and motion. Outputs up to 720p; the main model is about 80 GB.

CWCrisper Whisper 2Nyra LabsAvailable in EU

Open-source transcription model with verbatim and intended modes plus word-level timestamps. Creator tested it on a Hugging Face space; it beats ElevenLabs on its own benchmark. Four sizes from 0.2B to 2B parameters.

DVDeepSeek V4 FlashDeepSeekAvailable in EU

Updated open-source Flash model that scores about one point below GLM 5.2 and beats the previous V4 Pro. It is roughly 100x cheaper than Claude Opus at about 3 cents per million tokens. Weights are 167 GB, with Unsloth GGUFs down to 82.5 GB.

RERedesignAvailable in EU

Turns a flat image into editable layers, like a screenshot becoming a Figma or Photoshop project. It orchestrates PaddleOCR, Qwen Image Layered, Grounding DINO and SAM 2. Code is under 4 GB but needs an OpenAI API key by default.

KKKimi K3Moonshot AIUnverified

Moonshot released the open weights for its 2.8T-parameter MoE (104B active) with native vision, Kimi Delta Attention and attention residuals. Full weights are 1.56 TB, and the Unsloth 1-bit GGUF is 594 GB.

IMInstella MoEAMDAvailable in EU

Open-source 16B MoE (2.8B active) model trained from scratch on AMD Instinct GPUs with the ROCm stack. It reportedly beats similar-sized models such as Gemma 4 E4B, and AMD published checkpoints, recipes and code.

IOIdeogram Object RemoverIdeogramUnverified

Brush over an object and it is removed, including its shadows and reflections. It shows the lowest error rate on a removal benchmark versus Nano Banana 2 and GPT Image 2 Medium. Free to try with daily credits.

ISInkling SmallThinking MachinesAvailable in EU

Smaller open-source omnimodal model (text, audio, image, video) with 276B total and 12B active parameters, about a quarter the size of the full Inkling. It trails DeepSeek V4 Flash in intelligence but is stronger on audio. Weights are 532 GB.

SPSarvam PRISMSarvam AIAvailable in EU

Robot control system that turns sensor readings, images and instructions into movement actions, using combined signals such as force, velocity, contact and friction. It shows a higher success rate than comparable algorithms, and code is released.

S2Seedance 2.5ByteDanceUnverified

ByteDance's video model, which the creator calls the best available, with strong action scenes and character consistency. It generates up to 30 seconds at 720p and takes up to 50 multimodal references. A 10-second clip costs about $4.60 and the API is not out yet.

MHMiniMax H3MiniMaxUnverified

MiniMax's multimodal video model (successor to Hailuo) that takes text, image, video and audio references and generates 2K video up to 15 seconds. A 10-second clip costs about $1.20. An open-source release is expected next week.

GRGemini Robotics 2Google DeepMindUnverified

A family of three models for whole-body robot control: a vision-language-action model, ER 2 for high-level reasoning and planning, and On-Device 2 for offline use. It adds better dexterity, such as unscrewing a bulb or sealing a Ziploc bag. It can be tried in AI Studio.

WOWonderAdobeUnverified

Video world model that creates interactive worlds you navigate in real time with keys. It also accepts an input video to re-render with free camera movement. Quality shows noise and artifacts, and code and models are listed as coming soon.

GVGemini voice typingGoogleUnverified

Mac Gemini app feature: hold the function key and speak into any application. It transcribes with Gemini, removes filler words and self-corrections, fixes punctuation, and inserts text at the cursor. An optional reasoning mode can handle tasks like summarizing highlighted documents.

PHPhi-0Unverified

Video world model that first reasons about physical motion in a 'physical language' before rendering frames. It is applied to interactive worlds, autonomous driving and robot training videos, and it outperforms similar world models on physical coherence. Code is coming soon.