All Models

Ultra-tiny language model — runs anywhere

Ultra-small language model

Lightweight chat — surprisingly capable

Google's compact model — great reasoning for its size

Liquid AI's hybrid model — blazing fast CPU inference

Popular tiny model trained on 3T tokens

Meta's solid all-rounder

Stability AI's efficient small model

Largest SmolLM2 — punches above its weight

Google's on-device multimodal — 2B effective params with vision and audio

Sweet-spot model — great quality-to-size ratio

Meta's 3B — best small model for many tasks

Google's most capable on-device model — 4B effective, multimodal

Google's 4B — strong reasoning and instruction following

Top-tier 7B chat — rivals much larger models

Meta's workhorse 8B — excellent all-around

PrismML's native 1-bit model — 8B params in 1.16 GB, needs forked llama.cpp

PrismML's 1.58-bit ternary model — 1.7B params in under 0.4 GB, runs anywhere

PrismML's 1.58-bit ternary model — 4B-class intelligence in ~0.86 GB, ~9x smaller than fp16

PrismML's 1.58-bit ternary model — top intelligence at 8B in ~1.6 GB

Google's largest open model — near-frontier quality

Mistral's 12B — Tekken tokenizer, 128K context

Google's 2026 on-device model — 2.3B active / 5.1B total, multimodal vision and audio

Google's most capable edge model — 4B effective, multimodal, 128K context

Google's sparse MoE — 26B total, 3.8B active per token, 256K context

Google's dense flagship open model — near server-grade quality, 256K context

Alibaba's tiny multimodal model — vision-capable, 262K context, runs on phones

Compact multimodal chat — native vision, 262K context

Multimodal sweet spot — vision, strong reasoning, 262K context

Flagship small Qwen — native multimodal, 262K context, rivals far larger models

Alibaba's sparse MoE — 35B total, 3B active per token, multimodal, 262K context

Liquid AI's on-device MoE — 8.3B total, 1.5B active, reasoning + tool calling at 3–4B quality

Distilled reasoning — chain-of-thought in a tiny package

Microsoft's reasoning model — exceptional for its size

Distilled from DeepSeek R1 — strong step-by-step reasoning

High-quality reasoning and analysis

Llama-based R1 distill — best open reasoning at 8B

Alibaba's latest — thinking mode with strong reasoning

Microsoft's 14B reasoning model — frontier-class performance

Meta's MoE model — 109B total params, 17B active per token

DeepSeek's V4 Flash MoE — 284B total, 13B active, 1M context (server-grade hardware)

Moonshot's 1T MoE — 32B active, frontier agentic coding; needs a multi-GPU workstation

Tiny code completion model — autocomplete on any device

Best small code model — great for autocomplete

BigCode's multilingual code model — 600+ languages

Strong code generation and editing at 3B

DeepSeek's code model — strong at generation and debugging

Best open 7B code model — rivals GPT-4 on coding benchmarks

Poolside's open agentic coding MoE — 33B total, 3B active, runs locally on 36GB Macs

Alibaba's local coding agent MoE — 80B total, 3B active, agentic long-horizon coding, 256K context

Tiny vision-language model — describe images on any device

Small but capable vision model — image Q&A and captioning

Leading open vision model — image understanding and reasoning

State-of-the-art vision-language — image, video, document understanding

Fast sentence embeddings — ideal for semantic search

Compact BAAI embedding — great for RAG pipelines

Open-source embedding with 8K context — long document search

High-quality embeddings — top of MTEB benchmark at its size

1-step image generation — instant results

High-res 1-step generation — 1024×1024 in seconds

Latest SD architecture — excellent image quality

Black Forest Labs' fast model — stunning quality in 4 steps

PrismML's 1.58-bit ternary diffusion — first 4B image model to run on iPhone, ~3GB in-browser

Tiny, fast voice synthesis

24 natural voices — instant TTS

Voice cloning and natural TTS — zero-shot voice synthesis

Nari Labs dialogue TTS — multi-speaker with emotion

Neuphonic's on-device TTS — 748M, instant voice cloning, runs on CPU via llama.cpp

KittenML's expressive 80M TTS — 8 voices, quantization-aware, runs on any CPU

Real-time speech recognition — runs on anything

High-accuracy English ASR

Good accuracy-speed balance — 99 languages

Strong multilingual transcription

Best open transcription — near-human accuracy

6x faster than Whisper Large — nearly same accuracy