AI Model Explorer

Find the right AI model. Start with what you need.

Explore open models in plain English. Choose a task, understand the tradeoffs, and compare practical options.

This page helps you explore and compare models. It does not run them, download them, or upload anything you type.

How this works

Pick the task first. A model is a specialized tool. Choosing what you need done (transcribe, summarize, cut out a background) narrows things down faster than starting from a model name.
Access is not the same as permission. “Open weights” and “source-available” describe whether you can download a model. Its license decides what you are allowed to do with it, including commercial use.
Running a model locally takes more than a download. You still need a compatible app or runtime, the model files, and enough memory and processing power on your device.
A model card is not a hosted service. A listing here does not mean someone offers that model online, or that using it is free.
This is a small, curated collection. It is a starting point, not a live Hugging Face leaderboard or a list of the newest models.
Everything here happens in your browser. Filtering and comparison run on this page. Your searches and picks are not sent anywhere, no model runs, and no files are uploaded.

Not sure where to start? Help me choose Four quick questions, all optional

What do you need to do?

Write & summarize Remove a background Transcribe audio Write & explain code Understand an image Search by meaning Read text aloud Not sure yet

How do you want to run it?

You pick; this page does not inspect your device.

Only in a web browser, no installs I can install an app on a laptop I have a GPU or a server Not sure

Which language does it need to handle?

Will you use it commercially?

Yes, it must allow commercial use No, personal, study or research Not sure

How the shortlist is made

Task: only models for that task. Not sure: every task is considered.
Browser only: a browser-compatible conversion must be documented. Laptop: our laptop-friendly estimate, or a browser option. GPU: no limit. Setup labels are editorial estimates, not device tests.
Language: a match only when the model card lists it. A card that lists no languages, or an incomplete list, stays “unknown” and is never counted as a yes or a no.
Commercial use: “Yes” keeps only licenses this catalog marks as allowing commercial use. That is a reading of the license, not legal advice.
Results keep the catalog’s curated order. There are no scores, benchmarks or winners.

1. What do you want to do?

All models Every task, from text and code to images, audio, speech and search.

Pick a task above to compare its contrasting starting points.

2. Narrow it down

Setup labels are editorial starting points, not a promise that a model will run on your device. Curated order is a presentation order, and parameter count is model size, not a quality score.

3. Compare your options

Showing all 40 models

Writing Versatile writer

Qwen3 8B

Qwen

A general-purpose text assistant for drafting, rewriting, and working through questions.

Parameters 8.2B
GPU recommended

Good at: Flexible writing and conversation across many languages.
Keep in mind: Can invent facts; check names, numbers, and claims.

License Apache 2.0

Writing Compact text helper

SmolLM2 1.7B

Hugging Face

A smaller text assistant for straightforward rewrites, short summaries, and everyday prompts.

Parameters 1.7B
Laptop-friendly option
Browser option

Good at: Designed for on-device use with a smaller footprint.
Keep in mind: Less suited to difficult reasoning or complex instructions.

License Apache 2.0

Writing Tiny multilingual helper

Qwen3 0.6B

Qwen

A very small text assistant for short rewrites, quick drafts and simple questions in many languages.

Parameters 600M
Laptop-friendly option
Browser option

Good at: Supports 100+ languages and dialects, including translation.
Keep in mind: Its small size limits knowledge and accuracy; check every fact.

License Apache 2.0

Writing Open-weight reasoning

gpt-oss-20b

OpenAI

A larger open-weight model for reasoning-heavy writing, analysis and multi-step questions.

Parameters 21B
GPU recommended

Good at: Reasoning effort can be set to low, medium or high.
Keep in mind: Needs the harmony response format; it does not work correctly without it.

License Apache 2.0

Background removal Portrait specialist

MODNet

MODNet · Xenova conversion

Finds the person in a portrait so an app can separate them from the background.

Parameters Not listed
Laptop-friendly option
Browser option

Good at: Focused on human portraits and soft edges such as hair.
Keep in mind: Designed for people, rather than arbitrary product photos.

License Apache 2.0

Background removal General foregrounds

RMBG 2.0

BRIA

Separates foreground subjects from backgrounds across a wider range of image types.

Parameters Not listed
GPU recommended

Good at: Designed for varied subjects, including e-commerce imagery.
Keep in mind: Source-available weights are restricted to non-commercial use without an agreement.

License CC BY-NC 4.0

Commercial use restricted — check the license

Background removal High-detail cutouts

BiRefNet

ZhengPeng7

Separates the main subject from the background with fine edge detail across many kinds of images.

Parameters 220.7M
GPU recommended

Good at: Designed for high-resolution, fine-grained foreground segmentation.
Keep in mind: Heavier than the lite variant; slow without a GPU.

License MIT

Background removal Lighter general cutouts

BiRefNet Lite

ZhengPeng7

A smaller BiRefNet that removes backgrounds from general images with less compute.

Parameters 44.4M
Laptop-friendly option
Browser option

Good at: About a fifth of the full BiRefNet's size.
Keep in mind: Less precise on fine edges than the full model.

License MIT

Background removal Matting with edge refinement

BEN2

PramaLLC

Removes backgrounds from images and video, with a refiner step aimed at hair and fine edges.

Parameters 94.6M
Laptop-friendly option
Browser option

Good at: A refiner network reworks the pixels the base model is least sure about, such as hair.
Keep in mind: The open weights are the Base model; the publisher’s full model is offered only through its web demo and API.

License MIT

Background removal 2K-resolution cutouts

BiRefNet HR

ZhengPeng7

A BiRefNet trained on 2048x2048 images for cutouts from high-resolution photos.

Parameters 220.7M
GPU recommended

Good at: Trained at 2048x2048 for higher-resolution inference.
Keep in mind: At 1024x1024 the card’s scores are below the standard BiRefNet; use it for large images.

License MIT

Transcription Fuller transcription

Whisper Large v3 Turbo

OpenAI

Turns speech recordings into text in the spoken language.

Parameters 809M
GPU recommended

Good at: Multilingual speech recognition.
Keep in mind: Noisy audio, accents, and names can cause errors or invented words.

License MIT

Transcription Small & lightweight

Whisper Tiny

OpenAI

A compact speech-to-text model for quick experiments and short recordings.

Parameters 39M
Laptop-friendly option
Browser option

Good at: Much smaller than the larger Whisper variants.
Keep in mind: Expect more transcription errors than larger variants.

License Apache 2.0

Transcription Balanced transcription

Whisper Small

OpenAI

A mid-sized speech-to-text model: more accurate than Tiny, still light enough for a laptop.

Parameters 244M
Laptop-friendly option
Browser option

Good at: Noticeably stronger than Whisper Tiny for everyday recordings.
Keep in mind: Background noise and overlapping speakers still cause errors.

License Apache 2.0

Transcription Full-size accuracy

Whisper Large v3

OpenAI

The full-size Whisper model for the most accurate transcription and speech translation in the family.

Parameters 1.55B
GPU recommended

Good at: The card reports 10% to 20% fewer errors than Large v2 across many languages.
Keep in mind: Slow and memory-hungry without a GPU.

License Apache 2.0

Transcription Live transcription

Voxtral Mini 4B Realtime

Mistral AI

A streaming speech-to-text model that transcribes as audio arrives, with a delay you can set.

Parameters 4.4B
GPU recommended

Good at: Built for streaming: the card reports accuracy close to offline models at a 480 ms delay.
Keep in mind: Needs a GPU with at least 16GB of memory according to the card.

License Apache 2.0

Transcription Language ID + dialects

Qwen3-ASR 1.7B

Qwen

A speech-to-text model that also identifies the spoken language, including many Chinese dialects.

Parameters 1.7B
GPU recommended

Good at: Identifies the language and transcribes 30 languages and 22 Chinese dialects.
Keep in mind: Streaming works only with the vLLM backend, without timestamps or batching.

License Apache 2.0

Coding Coding specialist

Qwen2.5 Coder 7B

Qwen

A code-focused assistant for generating, explaining, and fixing snippets.

Parameters 7.61B
GPU recommended

Good at: Trained specifically for code generation and code reasoning.
Keep in mind: Generated code can contain bugs and insecure patterns.

License Apache 2.0

Coding Compact coding helper

Qwen2.5 Coder 1.5B

Qwen

A smaller member of the Qwen coding family for focused code tasks.

Parameters 1.54B
Laptop-friendly option

Good at: A lighter starting point for local coding experiments.
Keep in mind: Smaller capacity makes complex code changes harder.

License Apache 2.0

Coding Repository-scale coding

Qwen3 Coder 30B-A3B

Qwen

A large coding model for multi-file work, tool use and agent-style coding assistants.

Parameters 30.5B
GPU recommended

Good at: 262,144-token native context for repository-scale understanding.
Keep in mind: Needs substantial GPU memory even though few parameters are active.

License Apache 2.0

Coding Pocket-size code helper

Qwen2.5 Coder 0.5B

Qwen

The smallest Qwen coding model, for short snippets and explanations on almost any device.

Parameters 490M
Laptop-friendly option
Browser option

Good at: Small enough for browser experiments via an ONNX conversion.
Keep in mind: Struggles with multi-step logic and larger code changes.

License Apache 2.0

Coding Largest dense coder

Qwen2.5 Coder 32B

Qwen

The largest Qwen2.5 Coder model, for harder code generation, reasoning and fixing when you have a strong GPU.

Parameters 32.5B
GPU recommended

Good at: The card describes it as matching GPT-4o on coding at release.
Keep in mind: A dense 32.5B model: every parameter is used per token, so it needs far more GPU memory than the 7B variant.

License Apache 2.0

Coding Mid-size coder

Qwen2.5 Coder 14B

Qwen

A mid-size Qwen2.5 Coder model that sits between the 7B and 32B variants for code generation, reasoning and fixing.

Parameters 14.7B
GPU recommended

Good at: A step up from the 7B variant without the GPU memory needs of the 32B model.
Keep in mind: A dense 14.7B model: plan on a GPU with plenty of memory, or a quantized build.

License Apache 2.0

Image understanding Compact visual assistant

SmolVLM 500M

Hugging Face

Looks at an image and responds to a text question about it.

Parameters 500M
Laptop-friendly option

Good at: Small footprint for image-and-text experiments.
Keep in mind: Can miss small details or describe things that are not present.

License Apache 2.0

Image understanding Structured vision tasks

Florence-2 Base FT

Microsoft

Uses task prompts to describe images, read visible text, or locate objects.

Parameters 230M
Laptop-friendly option

Good at: Combines captioning, object detection, and OCR-style tasks.
Keep in mind: Requires specific task prompts and output parsing.

License MIT

Image understanding Documents, charts & video

Qwen2.5 VL 7B

Qwen

A capable vision-language model that reads documents, charts and long videos and answers questions about them.

Parameters 8.29B
GPU recommended

Good at: Can locate objects and return coordinates as JSON.
Keep in mind: Needs a GPU for comfortable speed, especially with video.

License Apache 2.0

Image understanding Small captions & queries

Moondream 2

Moondream (vikhyatk)

A small vision model that captions images, answers questions and points to objects.

Parameters 1.93B
Laptop-friendly option
Browser option

Good at: Built-in caption, query, detect and point functions.
Keep in mind: The publisher now points to Moondream 3 (Preview) as its successor.

License Apache 2.0

Image understanding Images, video & agents

Qwen3.8 27B

Qwen

A dense vision-language model that understands images and long videos and also handles coding and multi-step agent tasks.

Parameters 27.78B
GPU recommended

Good at: Native image and video understanding, from documents and STEM diagrams to hour-scale videos.
Keep in mind: Much heavier than the other vision options; plan on a strong GPU.

License Apache 2.0

Image understanding Multilingual OCR & UI

Gemma 4 31B

Google DeepMind

Google's largest dense Gemma 4 model, reading documents, charts, screenshots and handwriting, with broad multilingual support.

Parameters 30.7B
GPU recommended

Good at: Card lists document/PDF parsing, UI understanding, chart reading, multilingual OCR, handwriting recognition and pointing.
Keep in mind: The card targets consumer GPUs and workstations; it is not a laptop model.

License Apache 2.0

Semantic search Lightweight meaning search

MiniLM L6 v2

Sentence Transformers

Turns short text into numbers that help find passages with similar meaning.

Parameters 22.7M
Laptop-friendly option
Browser option

Good at: A compact choice for English sentence similarity.
Keep in mind: Does not write answers; an app must compare the vectors.

License Apache 2.0

Semantic search Search across languages

Multilingual E5 Small

intfloat

Encodes text for meaning-based search across multiple languages.

Parameters 118M
Laptop-friendly option

Good at: Supports 100 languages, with varying quality.
Keep in mind: Long text is truncated at 512 tokens.

License MIT

Semantic search Compact English retrieval

BGE Small EN v1.5

BAAI

Turns English text into compact vectors for search, matching and grouping by meaning.

Parameters 33.4M
Laptop-friendly option
Browser option

Good at: Small and fast, with 384-dimensional vectors.
Keep in mind: English only; inputs beyond 512 tokens are truncated.

License MIT

Semantic search Multilingual long documents

BGE-M3

BAAI

A multilingual embedding model for long documents that supports several retrieval styles at once.

Parameters Not listed
Laptop-friendly option
Browser option

Good at: Works across 100+ languages with inputs up to 8,192 tokens.
Keep in mind: Much larger than BGE Small; slower on modest hardware.

License MIT

Semantic search Instruction-aware search

Qwen3 Embedding 0.6B

Qwen

A compact multilingual embedding model with a long context window that you can steer with a task instruction.

Parameters 600M
Laptop-friendly option
Browser option

Good at: Supports 100+ languages, including programming languages, with inputs up to 32K tokens.
Keep in mind: The card reports instructions typically add 1% to 5%; it advises writing them in English.

License Apache 2.0

Semantic search Compressible vectors

Arctic Embed L v2.0

Snowflake

A multilingual retrieval embedding model built to keep English quality and to shrink its vectors for large indexes.

Parameters 568M
Laptop-friendly option
Browser option

Good at: Strong English and multilingual retrieval scores on the benchmarks the card reports.
Keep in mind: Queries need the “query: ” prefix; documents do not.

License Apache 2.0

Text to speech Tiny natural voices

Kokoro 82M

hexgrad

A very small text-to-speech model that reads text aloud in a choice of built-in voices.

Parameters 82M
Laptop-friendly option
Browser option

Good at: Only 82 million parameters, yet designed to sound comparable to larger models.
Keep in mind: Uses its built-in voices; it does not clone a voice from a sample.

License Apache 2.0

Text to speech CPU-friendly accents

MeloTTS English

MyShell

A text-to-speech model that reads English aloud with several accents and runs fast on a regular CPU.

Parameters Not listed
Laptop-friendly option

Good at: The card describes it as fast enough for real-time speech on a CPU.
Keep in mind: Fixed voices only; no voice cloning or emotion control.

License MIT

Text to speech Expressive voice cloning

Chatterbox

Resemble AI

An expressive text-to-speech model that can speak in a voice taken from a short reference recording.

Parameters 500M
GPU recommended

Good at: Zero-shot voice cloning plus an emotion-exaggeration control.
Keep in mind: Only clone a voice with that speaker's clear permission.

License MIT

Text to speech Instruction-controlled voices

Qwen3 TTS 1.7B CustomVoice

Qwen

A multilingual text-to-speech model whose tone, pace and emotion you can steer with plain-language instructions.

Parameters 1.7B
GPU recommended

Good at: Covers 10 major languages and offers 9 premium preset voices.
Keep in mind: Much larger than the lightweight voices; plan on a GPU.

License Apache 2.0

Text to speech Vietnamese + English voices

VieNeu TTS v3 Turbo

pnnbao-ump

A Vietnamese text-to-speech model with English code-switching, preset regional voices and voice cloning from a short clip.

Parameters 130.9M
Laptop-friendly option

Good at: 23 preset voices across North, Central and South Vietnamese, at 48 kHz.
Keep in mind: Built for Vietnamese and English only; emotion cues are marked experimental.

License Apache 2.0

Text to speech 30 languages + voice design

VoxCPM2

OpenBMB

A multilingual text-to-speech model that can design a new voice from a written description or clone one from a short clip, with 48 kHz output.

Parameters 2B
GPU recommended

Good at: Covers 30 languages with no language tag needed, plus several Chinese dialects.
Keep in mind: Needs an NVIDIA GPU with CUDA; the card lists about 8 GB of VRAM.

License Apache 2.0

Catalog snapshot: 2026-10-30. A small curated collection, not a live leaderboard.