---
title: "Model Registry | Xybrid"
description: "Browse 41 curated on-device AI models, or bring your own ONNX or GGUF model with Xybrid. Run locally on iOS, Android, macOS, and more."
url: "https://www.xybrid.ai/models"
---

# Find your model.

Text, voice, and vision. Start with the curated registry, or bring your own ONNX or GGUF model.

The model registry: 41 models. Sizes show the default model file; some variants need extra assets.

- [The same registry, as JSON](https://registry.xybrid.dev/v1/models)
- [Try a model in your browser](https://www.xybrid.ai/demo)

## Text generation (19)

| Model | ID | Lab | Parameters | Download | Format | Description |
| --- | --- | --- | --- | --- | --- | --- |
| [Functiongemma 270M It](https://www.xybrid.ai/models/functiongemma-270m-it) | `functiongemma-270m-it` | Google | 270M | 278.1 MB | GGUF | Google FunctionGemma 270M instruction-tuned function-calling model for local inference. |
| [Gemma 3 1B](https://www.xybrid.ai/models/gemma-3-1b) | `gemma-3-1b` | Google | 1B | 768.7 MB | GGUF | Gemma 3 1B - Google's mobile-optimized instruction-tuned LLM with 32K context |
| [Gemma 4 E2b](https://www.xybrid.ai/models/gemma-4-e2b) | `gemma-4-e2b` | Google | 5.1B | 2.9 GB | GGUF | Gemma 4 E2B - Google's compact multimodal LLM with 2.3B effective params, 128K context, and audio/image/video understanding |
| [Gemma 4 E4b](https://www.xybrid.ai/models/gemma-4-e4b) | `gemma-4-e4b` | Google | 8B | 4.6 GB | GGUF | Gemma 4 E4B - Google's mid-range multimodal LLM with 4.5B effective params, 128K context, and audio/image/video understanding |
| [Gemma3npc 1B](https://www.xybrid.ai/models/gemma3npc-1b) | `gemma3npc-1b` | Google | 1B | 768.7 MB | GGUF | Gemma3NPC 1B - Mobile-friendly NPC roleplay fine-tune of Gemma 3 1B for in-character game dialogue |
| [Gemma3npc It](https://www.xybrid.ai/models/gemma3npc-it) | `gemma3npc-it` | Google | 7B | 3.9 GB | GGUF | Gemma3NPC-it - NPC roleplay fine-tune of Gemma3n-E4B for in-character game dialogue |
| [Lfm2.5 1.2B Instruct](https://www.xybrid.ai/models/lfm2.5-1.2b-instruct) | `lfm2.5-1.2b-instruct` | Liquid AI | 1.2B | 697.0 MB | GGUF | Liquid LFM2.5 1.2B Instruct - hybrid conv+attention LLM optimized for edge deployment (agentic tasks, data extraction, RAG, tool calling) |
| [Lfm2.5 1.2B Thinking](https://www.xybrid.ai/models/lfm2.5-1.2b-thinking) | `lfm2.5-1.2b-thinking` | Liquid AI | 1.2B | 697.0 MB | GGUF | Liquid LFM2.5 1.2B Thinking - hybrid conv+attention reasoning LLM optimized for edge deployment (fast CPU inference, low memory) |
| [Lfm2.5 230M](https://www.xybrid.ai/models/lfm2.5-230m) | `lfm2.5-230m` | Liquid AI | 230M | 146.3 MB | GGUF | Liquid LFM2.5 230M - smallest hybrid conv+attention LLM optimized for edge deployment (9 languages, tool calling, fast CPU inference) |
| [Lfm2.5 350M](https://www.xybrid.ai/models/lfm2.5-350m) | `lfm2.5-350m` | Liquid AI | 354M | 218.7 MB | GGUF | Liquid LFM2.5 350M - hybrid conv+attention LLM optimized for edge deployment (9 languages, tool calling) |
| [Lfm3 1.7B](https://www.xybrid.ai/models/lfm3-1.7b) | `lfm3-1.7b` | Liquid AI | 1.7B | 1001.0 MB | GGUF | Liquid LFM3 1.7B - tri-hybrid conv+attention reasoning LLM for local agents and structured outputs (16 languages) |
| [Llama 3.2 1B](https://www.xybrid.ai/models/llama-3.2-1b) | `llama-3.2-1b` | Meta | 1B | 770.3 MB | GGUF | Llama 3.2 1B - Meta's lightweight mobile-optimized instruction-tuned LLM with 128K context |
| [Ministral 3 3B](https://www.xybrid.ai/models/ministral-3-3b) | `ministral-3-3b` | Mistral AI | 3.4B | 2.0 GB | GGUF | Ministral 3 3B Instruct - Mistral AI's edge-optimized instruction-tuned LLM with 256K context |
| [Mistral 7B](https://www.xybrid.ai/models/mistral-7b) | `mistral-7b` | Mistral AI | 7B | 4.1 GB | GGUF | Mistral 7B Instruct v0.3 - High-quality desktop LLM with function calling support |
| [Phi4 Mini](https://www.xybrid.ai/models/phi4-mini) | `phi4-mini` | Microsoft | 3.8B | 2.3 GB | GGUF | Phi-4 Mini 3.8B - Microsoft's compact reasoning LLM with 128K context |
| [Qwen2.5 0.5B Instruct](https://www.xybrid.ai/models/qwen2.5-0.5b-instruct) | `qwen2.5-0.5b-instruct` | Qwen | 500M | 468.6 MB | GGUF | Qwen 2.5 0.5B Instruct - Small but capable instruction-tuned LLM from Alibaba Cloud |
| [Qwen3.5 0.8B](https://www.xybrid.ai/models/qwen3.5-0.8b) | `qwen3.5-0.8b` | Qwen | 800M | 495.7 MB | GGUF | Qwen 3.5 0.8B - Alibaba Cloud's lightweight multimodal LLM (text-only mode, 201 languages) |
| [Qwen3.5 2B](https://www.xybrid.ai/models/qwen3.5-2b) | `qwen3.5-2b` | Qwen | 2B | 1.2 GB | GGUF | Qwen 3.5 2B - Alibaba Cloud's compact multimodal LLM (text-only mode, 201 languages) |
| [Smollm2 360M](https://www.xybrid.ai/models/smollm2-360m) | `smollm2-360m` | Hugging Face | 360M | 258.1 MB | GGUF | SmolLM2 360M - HuggingFace's best tiny LLM, excellent quality/size ratio |

## Text to speech (8)

| Model | ID | Lab | Parameters | Download | Format | Description |
| --- | --- | --- | --- | --- | --- | --- |
| [Kitten Tts Micro 0.8](https://www.xybrid.ai/models/kitten-tts-micro-0.8) | `kitten-tts-micro-0.8` | KittenML | 40M | 90.2 MB | ONNX | KittenTTS Micro 0.8 - Compact StyleTTS 2 (40M params, 8 named voices, OpenPhonemizer) |
| [Kitten Tts Mini 0.8](https://www.xybrid.ai/models/kitten-tts-mini-0.8) | `kitten-tts-mini-0.8` | KittenML | 80M | 111.1 MB | ONNX | KittenTTS Mini 0.8 - High-quality StyleTTS 2 (80M params, 8 named voices, OpenPhonemizer) |
| [Kitten Tts Nano 0.2](https://www.xybrid.ai/models/kitten-tts-nano-0.2) | `kitten-tts-nano-0.2` | KittenML | 15M | 18.3 MB | ONNX | KittenTTS Nano 0.2 - Ultra-lightweight TTS (15M params, <25MB) |
| [Kitten Tts Nano 0.8](https://www.xybrid.ai/models/kitten-tts-nano-0.8) | `kitten-tts-nano-0.8` | KittenML | 15M | 78.3 MB | ONNX | KittenTTS Nano 0.8 - Ultra-lightweight StyleTTS 2 (15M params, 8 named voices, OpenPhonemizer) |
| [Kokoro 82M](https://www.xybrid.ai/models/kokoro-82m) | `kokoro-82m` | Hexgrad | 82M | 174.9 MB | ONNX | Kokoro 82M - High-quality TTS with 24 voices, Misaki dictionary |
| [Neutts Air Q4](https://www.xybrid.ai/models/neutts-air-q4) | `neutts-air-q4` | Neuphonic | 500M | 755.0 MB | GGUF | NeuTTS Air (~500M) — codec TTS with voice cloning. Q4 GGUF backbone + NeuCodec decoder. Higher quality than Nano. |
| [Neutts Nano Q4](https://www.xybrid.ai/models/neutts-nano-q4) | `neutts-nano-q4` | Neuphonic | 120M | 469.7 MB | GGUF | NeuTTS Nano (120M) — codec TTS with voice cloning. Q4 GGUF backbone + NeuCodec ONNX decoder. |
| [Phonon 31M En](https://www.xybrid.ai/models/phonon-31m-en) | `phonon-31m-en` | Gradium | 31M | 71.0 MB | GGUF | Phonon 31M - English streaming TTS from Gradium (24 kHz, 38 built-in voices), one GGUF for the zzz CPU engine |

## Speech to text (9)

| Model | ID | Lab | Parameters | Download | Format | Description |
| --- | --- | --- | --- | --- | --- | --- |
| [Wav2vec2 Base 960H](https://www.xybrid.ai/models/wav2vec2-base-960h) | `wav2vec2-base-960h` | Meta | 95M | 220.3 MB | ONNX | Wav2Vec2 Base 960h - English ASR with CTC decoding |
| [Whisper Base En Ggml](https://www.xybrid.ai/models/whisper-base-en-ggml) | `whisper-base-en-ggml` | OpenAI | 74M | 57.0 MB | GGML | Whisper Base English - English-only ASR (whisper.cpp/GGML runtime) |
| [Whisper Base Ggml](https://www.xybrid.ai/models/whisper-base-ggml) | `whisper-base-ggml` | OpenAI | 74M | 56.9 MB | GGML | Whisper Base - Multilingual ASR (whisper.cpp/GGML runtime) |
| [Whisper Large V3 Turbo Ggml](https://www.xybrid.ai/models/whisper-large-v3-turbo-ggml) | `whisper-large-v3-turbo-ggml` | OpenAI | 809M | 547.4 MB | GGML | Whisper Large v3 Turbo - Highest-accuracy multilingual ASR (whisper.cpp/GGML runtime) |
| [Whisper Medium Ggml](https://www.xybrid.ai/models/whisper-medium-ggml) | `whisper-medium-ggml` | OpenAI | 769M | 514.2 MB | GGML | Whisper Medium - Multilingual ASR (whisper.cpp/GGML runtime) |
| [Whisper Small Ggml](https://www.xybrid.ai/models/whisper-small-ggml) | `whisper-small-ggml` | OpenAI | 244M | 181.3 MB | GGML | Whisper Small - Multilingual ASR (whisper.cpp/GGML runtime) |
| [Whisper Tiny](https://www.xybrid.ai/models/whisper-tiny) | `whisper-tiny` | OpenAI | 39M | 84.7 MB | SAFETENSORS | Whisper Tiny - Fast multilingual ASR (Candle/SafeTensors runtime) |
| [Whisper Tiny En Ggml](https://www.xybrid.ai/models/whisper-tiny-en-ggml) | `whisper-tiny-en-ggml` | OpenAI | 39M | 30.7 MB | GGML | Whisper Tiny English - Fastest English-only ASR (whisper.cpp/GGML runtime) |
| [Whisper Tiny Ggml](https://www.xybrid.ai/models/whisper-tiny-ggml) | `whisper-tiny-ggml` | OpenAI | 39M | 30.7 MB | GGML | Whisper Tiny - Fast multilingual ASR (whisper.cpp/GGML runtime) |

## Vision + language (4)

| Model | ID | Lab | Parameters | Download | Format | Description |
| --- | --- | --- | --- | --- | --- | --- |
| [Bonsai 27B](https://www.xybrid.ai/models/bonsai-27b) | `bonsai-27b` | PrismML | 27.3B | 3.5 GB | GGUF | PrismML Bonsai 27B 1-bit vision-language model for local reasoning. |
| [Lfm2 Vl 450M](https://www.xybrid.ai/models/lfm2-vl-450m) | `lfm2-vl-450m` | Liquid AI | 450M | 209.1 MB | GGUF | Liquid LFM2-VL 450M compact vision-language model for edge inference. |
| [Lfm2.5 Vl 3B](https://www.xybrid.ai/models/lfm2.5-vl-3b) | `lfm2.5-vl-3b` | Liquid AI | 3B | 1.6 GB | GGUF | Liquid LFM2.5-VL 3B vision-language model for local inference. |
| [Lfm3 Vl 2B](https://www.xybrid.ai/models/lfm3-vl-2b) | `lfm3-vl-2b` | Liquid AI | 1.9B | 1001.0 MB | GGUF | Liquid LFM3-VL 2B vision-language model for on-device visual tasks: documents, charts, app screens and object grounding. |

## Decision (1)

| Model | ID | Lab | Parameters | Download | Format | Description |
| --- | --- | --- | --- | --- | --- | --- |
| [D1 2B](https://www.xybrid.ai/models/d1-2b) | `d1-2b` | Liquid AI | 1.7B | 1.0 GB | GGUF | Liquid d1-2B - decision model: typed choice, yes/no and score answers with probabilities over text or JSON, no generated text (16 languages) |

## Bring your own model.

Use any ONNX or GGUF model compatible with your runtime. Describe its files and execution pipeline in model_metadata.json, check device resources, then run it locally. [Prepare your model](/bring-your-own-model.md)
