Models

Every model in the Xybrid registry runs fully on-device — no cloud account or API key required. Models are downloaded once, cached locally, and resolved automatically to the best variant for your platform.

The registry is the source of truth. The tables below are a snapshot — run xybrid models list (or xybrid models search <query>) for the live catalog.

Start with these

Model	Task	Params	Why start here
`smollm2-360m`	LLM	360M	Best quality-to-size ratio for any device
`kokoro-82m`	TTS	82M	High-quality speech, 24 voices, fast
`whisper-tiny`	ASR	39M	Accurate multilingual transcription

Text generation (LLM)

LLMs ship as Q4_K_M-quantized GGUF and run on the llama.cpp backend.

Model ID	Params	Download	Description
`smollm2-360m`	360M	267 MB	HuggingFace's best tiny LLM, excellent quality/size ratio
`qwen2.5-0.5b-instruct`	500M	477 MB	Small but capable instruction-tuned LLM
`lfm2.5-350m`	354M	—	Liquid LFM2.5, hybrid conv+attention LLM for edge
`qwen3.5-0.8b`	800M	520 MB	Lightweight multimodal LLM (text-only mode)
`gemma-3-1b`	1B	786 MB	Google's mobile-optimized LLM with 32K context
`gemma3npc-1b`	1B	845 MB	NPC roleplay fine-tune of Gemma 3 1B
`llama-3.2-1b`	1B	791 MB	Meta's lightweight mobile-optimized LLM
`qwen3.5-2b`	2B	1.3 GB	Compact multimodal LLM (text-only mode)
`ministral-3-3b`	3.4B	2.1 GB	Mistral AI's edge-optimized instruction model
`phi4-mini`	3.8B	8.7 GB	Microsoft's compact reasoning LLM, 128K context
`gemma-4-e2b`	5.1B	—	Google's compact multimodal LLM (2.3B effective)
`gemma3npc-it`	7B	4.1 GB	NPC roleplay fine-tune of Gemma3n-E4B
`mistral-7b`	7B	4.3 GB	High-quality desktop LLM with function calling
`gemma-4-e4b`	8B	—	Google's mid-range multimodal LLM (4.5B effective)

Text-to-speech (TTS)

Model ID	Params	Download	Description
`kitten-tts-nano-0.2`	15M	19 MB	Ultra-lightweight TTS (under 25 MB)
`kitten-tts-nano-0.8`	15M	82 MB	Ultra-lightweight StyleTTS 2, 8 named voices
`kitten-tts-micro-0.8`	40M	95 MB	Compact StyleTTS 2, 8 named voices
`kitten-tts-mini-0.8`	80M	116 MB	High-quality StyleTTS 2, 8 named voices
`kokoro-82m`	82M	183 MB	High-quality TTS with 24 voices
`neutts-nano-q4`	120M	493 MB	Codec TTS with voice cloning (Q4 GGUF backbone)
`neutts-air-q4`	~500M	792 MB	Codec TTS with voice cloning (Q4 GGUF backbone)

List the voices of any TTS model with xybrid models voices <model-id>.

Speech-to-text (ASR)

Model ID	Params	Download	Description
`whisper-tiny`	39M	89 MB	Fast multilingual ASR (Candle/SafeTensors runtime)
`wav2vec2-base-960h`	95M	231 MB	English ASR with CTC decoding

Model naming

Model IDs follow the pattern {family}-{variant}-{version}:

kitten-tts-nano-0.2 — family, variant, and upstream model version
kokoro-82m — no version suffix when only one release exists
The version is the upstream model version, not Xybrid's integration version

Registry entries are immutable once published — a new upstream version becomes a new registry entry.

Using a model

# CLI
xybrid models info kokoro-82m
xybrid run --model kokoro-82m --input-text "Hello!" -o hello.wav

// Flutter — same model IDs across every SDK
final model = await Xybrid.model('kokoro-82m').load();

Downloads are cached at ~/.xybrid/cache/ — see Model Caching.

Models outside the registry

You can run models that aren't in the registry:

HuggingFace GGUF: xybrid run --huggingface <repo>[:variant] (or fromHuggingFace in the SDKs) downloads a GGUF directly
Local bundles: load a .xyb bundle or model directory
Custom models: prepare your own model_metadata.json — see the custom models guide