New releases

Gradium · Gradium's suite of on-device text-to-speech models under 50M parameters, built to run offline on phones, laptops, small boards and in the browser

Kyutai · Kyutai's Pocket TTS v4.0.0 release: six small English text-to-speech models with voice cloning, at 18.2M, 6.9M and 1.9M parameters, each with a…

NVIDIA · NVIDIA and the Baker Lab's flow-matching model for biomolecular design

Microsoft · Microsoft's 7B open-weights model, fine-tuned from Mistral-7B-Instruct-v0.2, that labels a user message about a service, product or agent as Negative, Neutral…

SenseTime · SenseTime's 27B embodied model for robot planning and decision-making, trained from Qwen3.8-27B with experience from its open-source SenseNova-RoboRSI harness…

Cloudflare · Cloudflare's System One extraction model, which extracts, computes, estimates and forecasts values from unstructured inputs

GoTo (with AI Singapore) · GoTo's 27B Gemma 3 instruct model for English and Indonesian regional languages (Javanese, Sundanese, Batak Toba, Balinese), developed with AI Singapore

Ant Ling (InclusionAI, Ant Group) · InclusionAI's open Ming-3.0 audio family: multilingual TTS with voice cloning, Chinese/English ASR with dialects, and native audio-language MoE models for…

Dream · Dream's open cybersecurity language models (8B and 30B-A3B), with a dedicated security mid-training stage before chat tuning

Mind Lab (Macaron AI) · Mind Lab's System 1 spatial decision model on Qwen3.5-4B: scores answer options, value bins and image regions with native attention Q/K, without autoregressive…

Ropedia · Ropedia's egocentric world foundation model trained on about 10k hours of first-person video for loco-manipulation that transfers across robot embodiments

Abacus.AI · Abacus.AI's open-weights detector for AI-generated images, released on Hugging Face under Apache 2.0 with GGUF builds

VAST · VAST's real-time world-model renderer that turns a rough game or reference video into detailed frames while upstream code keeps the world state

Blockway · Blockway's open-weights 32B model for agentic coding, the full release after its Preview, tuned for long agent sessions

OrcaRouter · A 27B open-weight vision-capable model built for defensive cybersecurity work such as vulnerability research and incident response

Birla AI Labs (Aditya Birla Group) · Birla AI Labs' first time-series foundation model: a 271.8M-parameter model for zero-shot probabilistic forecasting that returns a range of likely values…

Kyutai · Kyutai's 2.4B vision-language-action robot model, built with Inria and École des Ponts, pretrained on public data only, with a built-in memory of its recent…

NVIDIA · NVIDIA's 4B looped model: the same weights can run 1, 2 or 4 loops per token, trading compute for accuracy

Tencent · Tencent Hunyuan's dedicated long-term memory writer for AI agents, part of Hy Memory 2.0 alongside a framework and MemHub

OpenMOSS / MOSI · OpenMOSS / MOSI's text-to-image Diffusion Transformer, released as MOSS-Image-Base and MOSS-Image-SFT, with Chinese and English text rendering and controllable…

StartLux · StartLux's open-weight decision-model family: an Instant line from 0.8B to 27B and an AutoThink adaptive-reasoning line from 4B to 27B

Radical Numerics · Radical Numerics' family of early-fusion autoregressive text-to-image base models, trained entirely from scratch on 100M image-text pairs (GPIC) with no…

Skywork AI · Skywork's real-time video world model for games: a game engine keeps the world state and rules, and the video model renders each frame

NABLAS Inc. · NABLAS's Japanese fact-checking model (Qwen3-8B fine-tune) plus a citing agent that surfaces primary sources with the verdict; UTokyo spin-off work from GENIAC…