Local AI models, explained

Idol turns one photo, render or generated image of an object into a 3D Gaussian splat file.

Describe a scene or attach a picture and get a single web page that draws it in interactive 3D.

LTX-2.5 is a 22-billion-parameter open-weights model from Lightricks that generates video and sound together: describe a scene in words, or hand it a photo, and it produces a clip with a soundtrack in one pass, so picture and audio stay in sync.

Run a 27-billion-parameter model that reads text, images and video and writes text back, on hardware you control.

Transcribe recordings up to an hour long in 40 languages, with speaker labels and timestamps, or translate speech.

Create images from text or edit pictures using reference images inside ComfyUI, for people who make images and run ComfyUI themselves.

Turn short text requests into 8-megapixel images and edit pictures on your own graphics card.

Use its Python encoder to format text and image chats, then run the model and get written replies.

Chat with a small text model that can call tools and give structured answers, for developers building local agents or specialized apps.

Give a model up to a million tokens of text and images in one go, such as a long agent session or a large pile of input, and get text back.

Viggle’s turbo build is a distilled, faster version of Qwen-Image-2.1, the open image model: where the original takes 40 passes to render a picture, this one takes about six, which Viggle reports makes it roughly five times faster from prompt to finished image.

Unsloth’s release repackages Qwen-Image-2.1 — the open image-generation model — into GGUF files, the compact format that runs on ordinary computers through tools like llama.cpp and Stable Diffusion.cpp instead of demanding a datacenter GPU.

Ask questions about text, images and documents, write code, and get replies in a fixed data format.

Convert page images of long documents into text, with all pages read together in one request.

Run an open model that writes code and works through multi-step office tasks using tools.

For developers building voice apps: LFM3-Omni-2B answers questions about photos, video and audio, and replies in text, speech or both.

For developers building voice apps: LFM3-Audio-2B turns speech into text, text into speech, and answers spoken questions in text or voice.

Let an AI model investigate code, fix bugs, add features and write tests, for developers who host their own coding agents.

Run a compact language model on your own device for coding, tool-using agents and long inputs; for developers building on-device AI apps.

Send a screenshot, chart or document and get a written explanation, or hand it a bug to fix.

Ask a model questions about text or images and read written answers, through the Z.ai API or on your own servers.

Ask questions about document images, pictures and video and get text answers; it needs NVIDIA GPU servers, so it suits developers.

Send text, images or video and get written answers.

Generate 4–15 second videos with built-in stereo sound from text, images, video or audio, through an online API, web app or desktop app.

Create or edit pictures from written requests with this open weights image model, for developers building image tools.

Turn photos, scans and PDFs of Chinese or English documents into editable text, tables and formulas.

Turn a written prompt or a JSON layout into an image, with control over where text and objects go.

Turn text into multi-speaker speech with voice cloning and continuation controls.

A 740-million-parameter embedding model from Google DeepMind that turns text, code, images, video and audio into vectors in one shared 768-dimensional space.

Forecast future values from numeric time-series data, including covariates and uncertainty ranges.

Turn scanned PDFs and images into searchable text, Markdown, or JSON with boxes.

Generate images with planned text and object placement using this model for image makers, from written prompts or structured layouts.

Paste an AI-written email, essay or report and get a rewrite that reads more naturally, offline; for people writing in English or Chinese.

Ask questions about text, images, or video and receive text responses.

Extract and update conversation facts for memory frameworks such as Mem0 or MemOS.

Run an AI that thinks step by step and calls tools, on a laptop or your own server.

An accelerated 7-billion-parameter image model for making pictures from text and editing an existing image.

Turn photos of documents, multi-page scans and PDFs into text you can copy and search.

An experimental preview of the architecture Alibaba intends to build its next model generation on, released with open weights.

Ask questions about one image, several images or a video and get text answers.

Ask yes-or-no, multiple-choice or rating questions about a piece of text, JSON, an image or a video, and get a probability for every allowed answer.

A small model built to make decisions rather than write text.

JEV-35B evaluates up to 256 options in one pass.

Qwen-Image-2.1 is a 7-billion-parameter image model from Alibaba's Qwen team that both generates pictures from a text description and edits existing ones in the same model, using up to 10 reference images and letting you mark local edits with circles, painted annotations or masks while preserving the identity of people and products shown.

ZDTaichu5.0-9B is a 9-billion-parameter vision-language model, meaning it reads images, video and text together.

CLM-v0.1-8B is an 8-billion-parameter model that scores choices, such as which action an AI agent should take next, instead of writing text.

Julia-1 is a 144-million-parameter decision model, tiny beside the big chat models, that takes background context, a question and a list of possible answers, then picks the right one.

Nemotron-3-Diarization is an open-weight model from NVIDIA that does speaker diarization — working out "who spoke when" in a recording — for up to eight speakers.

Run a 27-billion-parameter reasoning model from a roughly 6 GB file on a laptop or a single GPU.

Run Qwen3.8-27B, a 27-billion-parameter model that reads images and text, on your own computer at a file size you choose.

TeleOCR is a vision-language model of about 1.2 billion parameters (learned settings) for reading documents.

Audio8 ASR Infinite is a speech-to-text model built for real-time, unlimited-length transcription rather than processing a finished audio file after the fact.

This repository repackages Alibaba's Qwen-Image-2.1 image generation and editing model as single-file weights formatted specifically for ComfyUI, a popular node-based visual tool for building AI image and video pipelines by connecting boxes on a canvas rather than writing code.

Run a language model that plans several steps, calls tools and works through long coding or research tasks, using a context of 256,000 tokens.

Hemmingway-1 is a 27-billion-parameter language model (built on Qwen3.5) fine-tuned specifically for the everyday writing people actually do — a text to a landlord, a work email, an awkward note to a colleague — rather than the longer, hedged answers general chat models tend to give.

MiMo-V2.6-Pro-RL is Xiaomi's flagship model in the MiMo-V2.6 family, reading and generating text, images, video and audio in one model with a 1-million-token context for handling large codebases or long multi-session agent work.

MiMo-V2.6-Distill-Qwen-9B is a 9-billion-parameter model made by fine-tuning Qwen3.5-9B on data generated by Xiaomi's larger MiMo-V2.6 models, a technique called distillation where a smaller model is trained to imitate a bigger one's outputs rather than being trained on raw internet text from scratch.

Give one model text, images, video or audio, with up to a million tokens of context such as a whole code repository or a long tool log, and get text back for coding and agent work.

A cross-encoder — a model that reads two pieces of text together and judges their relationship — built by further training Qwen3.5.

Sort a document into fixed answers, such as a priority level or a yes-or-no flag, and get back valid JSON in one quick step on an Apple Silicon Mac.

Get answers from a 27-billion-parameter reasoning model that spends less time thinking to itself before replying.

Compressed builds of Qwen3.8-27B, Alibaba’s 27-billion-parameter model that reads images and video as well as text, packaged as GGUF files so they run through llama.cpp, Ollama or LM Studio on hardware that could not hold the original.

A small language model — 2.5 billion parameters in a standard Llama-style architecture — built specifically for running on a device rather than in a data centre, with a 131,072-token context that is unusual at this size.

A compressed, ready-to-run version of Swift-Qwen3.8-27B, which is Qwen3.8-27B adjusted to reach its answers with far less internal deliberation.

The Apple Silicon build of a heavily compressed 27-billion-parameter model, storing each weight as one of three values — minus one, zero or plus one — in an 8.6 GB download that includes a vision component.