OpenAI's general-purpose speech recognition model, supporting multilingual transcription, translation, language identification, and multiple model sizes for local inference.openai
Microsoft's open-source family of voice AI models, including long-form speech recognition, multi-speaker TTS, and real-time streaming TTS weights.microsoft
An open-source foundation model for financial candlestick sequences, with pretrained tokenizers and models on Hugging Face plus forecasting and fine-tuning pipelines.shiyu-coder
Surya is a compact document OCR and layout vision-language model that recognizes text, layout, reading order and tables in 90+ languages, with a CLI and Python client for local inference.datalab-to
Chinese-extended LLaMA and instruction-tuned Alpaca large language models, releasing LoRA weights plus training and local quantized deployment guides.ymcui
Kitten-TTS is a family of compact open-source text-to-speech models, from 4 MB to 1.7B parameters, supporting multilingual speech, voice cloning, and expressive control on CPUs and consumer GPUs.kittenml
An open-source music generation foundation model that turns text into full songs and supports covers, repainting and vocal-to-BGM, running locally on consumer GPUs with LoRA personalization.ace-step
A frontier open music-generation model that plans melody and chords symbolically before rendering full songs with vocals, enabling zero-shot covers and agent-driven score editing.multimodal-art-projection
Kimi K2 is Moonshot AI's open-source family of trillion-parameter MoE language models, released as base and instruct weights for agentic, tool-use, and coding workloads. It can be deployed with vLLM, SGLang, or other inference engines.moonshotai
The official Python inference and LoRA training package for LTX-2, an audio-video generative model. It provides distilled, production-quality, image-to-video, keyframe, dubbing and HDR pipelines with memory and speed optimizations.lightricks
A family of small Qwen-based decision models for yes/no, multiple-choice and rating questions, with published weights, fine-tuning code and a local serving API.jaredpalmer
Kimi K3 is an open-weight 2.8T-parameter native multimodal MoE model with a one-million-token context window, built for long-horizon coding, knowledge work, and reasoning.moonshotai
Dolphin is ByteDance's open vision-language model for document image parsing. Its two-stage analyze-then-parse design handles digital and photographed pages, recovering layout, reading order, tables, formulas and code from PDFs and images.bytedance
A family of open LLMs dedicated to decompiling binaries into readable C code, with benchmarks, training scripts, and Hugging Face checkpoints.albertan017
SenseNova-U is a family of native multimodal models for image generation, editing, interleaved text-image output and visual understanding, released with inference examples, training code and weights.opensensenova
An open-source family of on-device text-to-speech models with instant voice cloning, multilingual support, and GGUF quantizations for running natural speech locally on phones, laptops, and embedded devices.neuphonic
A lightweight block-diffusion draft model for speculative decoding, shipped with checkpoints and tooling that plugs into Transformers, MLX, SGLang and vLLM to speed up LLM generation.z-lab
GPA is a unified autoregressive audio model for speech recognition, text-to-speech and voice conversion, released with 0.6B checkpoints, PyTorch/ONNX inference and fine-tuning pipelines.autoark
UniMate is a unified text-to-motion model that animates diverse skeletons without per-rig retraining. The release includes checkpoints, the UniML3D dataset, training and inference code, and tools for motion in-betweening, editing and expansion.friedrich-m
Qwen-UI-Agent (MAI-UI) is a foundation GUI agent model family for mobile, desktop, browser, and deep-search tasks. The repository releases model weights, benchmarks, and research results for real-world GUI automation.tongyi-mai
MMAudio is a CVPR 2025 model for high-quality video-to-audio synthesis, generating synchronized sound from video and text with released checkpoints and code.hkchengrex
An open 0.9B end-to-end audio model for long-form transcription in 50+ languages, with speaker diarization, timestamps, and acoustic event awareness. The repo ships inference code, serving recipes, a subtitle web app, and fine-tuning guidance.OpenMOSS
HunyuanOCR is a lightweight end-to-end OCR vision-language model unifying document parsing, text spotting, information extraction and text-image translation, released with training and inference code.tencent-hunyuan
JD's research release for long-horizon audio-visual generation, bundling the Echo-LongVideo model and the interactive Echo-WM world model with inference code and separately downloaded checkpoints.jd-opensource
FireRedAudio is a general-purpose audio language model release with inference code for speech recognition, understanding, voice cloning, TTS, and editing.fireredteam
LingBot-World 2.0 is an interactive world model release with 14B and 1.3B weights and inference code for long-horizon, action-driven video generation.robbyant
Descript's neural audio codec model, released with weights plus training and inference scripts. It compresses 44.1 kHz audio at very low bitrates while preserving fidelity.descriptinc
Google's open-weights model for real-time music generation, shipped with a Python inference library, a C++ streaming engine plus example apps and DAW plugins.magenta
The official implementation of RAVE, a realtime neural audio variational autoencoder. It includes training, export, and realtime inference tools for timbre transfer and music generation.acids-ircam
Tencent's family of universal multimodal embedding models, released in 2B, 4B and 9B sizes, mapping text, images, videos and visual documents into a shared retrieval space.tencent
Lightweight multimodal model family covering image and video understanding alongside text-to-image generation, with released weights supporting research, fine-tuning, and deployment by developers.microsoft
FireRedTTS3 is a speech generation and editing model release supporting zero-shot cloning in 24 languages and 21 Chinese dialects, plus voice design.fireredteam
ByteDance's family of multimodal large language models for speech, audio and audio-visual understanding, released with model checkpoints and inference and training code.bytedance
A multilingual speech recognition model for verbatim and intended transcription with precise word-level timestamps, distributed as model weights plus a Python inference package.nyrahealth
A 2B-parameter fully continuous end-to-end autoregressive TTS model with multilingual zero-shot voice cloning and precise speech editing, released with checkpoints, inference and fine-tuning code.rednote-hilab
ByteDance's unified video generation and editing framework pairing an MLLM semantic planner with a DiT renderer, released with inference code, training code, and model weights.bytedance
The official training and inference codebase for the WALL series of embodied foundation models, including LeRobot data prep, flow-matching/FAST action branches, serving and evaluation utilities.x-square-robot
A continually learning byte-level language model that trains from scratch on an 8 GB GPU, pages expert weights from disk and keeps learning from everything it reads.volotat
An open 9B vision model from Datalab that extracts schema-conformant JSON from PDFs and images, with local HuggingFace or vLLM-based deployment options.datalab-to
A family of time-series foundation models for multivariate probabilistic forecasting with lightweight inference code and leading zero-shot benchmark accuracy.mistralai
An open roadmap project for building world models to evaluate robot policies, sharing training and inference pipelines, model weights, datasets, and evaluation leaderboards.open-gigaai
A vision-language-action foundation model for real-world robot manipulation, released with pre-trained weights, training recipes and deployment code covering multiple embodiments and predictive dynamics.robbyant
Official research code for the paper Rethinking On-Policy Distillation of Large Language Models. It provides OPD, SFT and GRPO training recipes, evaluation scripts, released checkpoints and diagnostics for reproducing the reported findings.thunlp
An open family of medical vision-language models (4B–235B) for text, 2D/3D imaging and surgical video, released with inference and training code and strong benchmark results.zjui-ai4h
An open-source family of image generation and editing models, including Base, Turbo, and Edit variants for text-to-image, image editing, and bilingual text rendering, released with inference code and weights.boogu-project
An open-source mixture-of-experts video generation model for embodied intelligence, shipping weights, inference scripts, and benchmarks for text-to-video tasks.robbyant
The official implementation of EVOKE, an interactive world model that generates long, camera-controlled video with bounded state, released with inference code, checkpoints, and a director UI.alayalab
Tencent Hunyuan's open embodied vision-language foundation model for physical-world agents, with roughly 30B total and 3B active parameters, released with weights and Transformers/vLLM inference code.tencent-hunyuan
OpenCUA is an open foundation suite for computer-use agents, offering the OpenCUA vision-language model family, the AgentNet demonstration dataset, an annotation tool, and an offline action-level benchmark.xlang-ai
Confucius4-TTS is an open multilingual, cross-lingual zero-shot text-to-speech model that clones a speaker's voice across 14 languages while preserving emotion and identity.netease-youdao
An open-source interactive long-horizon video world model with real-time camera control, prompt switching, and long-horizon memory, released with training/inference code and staged model weights.alayalab
A general-purpose interactive world model that generates long-horizon, explorable video worlds controlled by camera motion and event prompts, released with 5B checkpoints and inference code.amap-ml
An open-source audio understanding model family for speech, environmental sound and music. The release provides 4B and 8B Instruct and Thinking weights for captioning, timestamped ASR, audio QA and reasoning.OpenMOSS
HelixWorld is a real-time interactive audio-visual world model from Noiz AI that generates synchronized visuals and camera-aligned spatial audio as you move, released with inference code and checkpoints.noizai
A 2.5B diffusion-based document OCR model that replaces autoregressive decoding with block-level parallel decoding. It ships model weights and inference scripts for layout, text, formula and table recognition.opendatalab
Liquid AI's LFM2-Audio-1.5B end-to-end speech-to-speech foundation model, shipped with Python inference, demo and fine-tuning code for real-time conversation, ASR and TTS.Liquid4All
A real-time multiplayer world model of Rocket League, releasing training code, datasets, and evaluation tools for interactive video generation research.mira-wm
Ant Group's LLaDA2.X series of diffusion language models, releasing 16B and 100B MoE weights with training code and an optimized dInfer/SGLang inference stack.inclusionai