Qwen3 0.6B
0.6B · 1.4GB · Apache-2.0
A hummingbird: instant answers on phones, Raspberry Pis and anything with a pulse.
Model directory
Every model your machine can run
Filter by what you actually own. VRAM numbers assume the community-standard Q4_K_M quantization; context and licenses included.
Qwen3 1.7B
1.7B · 1.9GB · Apache-2.0
Pocket assistant that still holds a coherent conversation.
Qwen3 4B
4B · 3.2GB · Apache-2.0
The sweet spot for laptops without dedicated GPUs.
Qwen3 8B ★
8B · 5.8GB · Apache-2.0
The default recommendation for a 12GB GPU: strong chat, solid code, hybrid thinking mode.
Qwen3 14B
14B · 10GB · Apache-2.0
16GB-card favorite with a clear step up in reasoning depth.
Qwen3 32B
32B · 20GB · Apache-2.0
Dense flagship for 24GB rigs — near-frontier quality, fully offline.
Qwen3 30B-A3B ★
30B · 19GB · Apache-2.0
Mixture-of-Experts trick: 30B knowledge with 3B speed. The people's flagship.
Qwen3 235B-A22B
235B · 135GB · Apache-2.0
Open MoE at frontier scale — multi-GPU or Mac Studio territory.
Llama 3.2 1B
1B · 1.6GB · Llama 3.2 Community
Meta's tiny edge model with a huge 128K context.
Llama 3.2 3B
3B · 2.7GB · Llama 3.2 Community
Balanced edge model; great multilingual chit-chat for its size.
Llama 3.1 8B ★
8B · 5.8GB · Llama 3.1 Community
The model that made local AI mainstream. Still a dependable daily driver.
Llama 3.3 70B
70B · 43GB · Llama 3.3 Community
Text-only 70B that punches like a 405B — 24GB+ dual-GPU class.
Llama 3.2 11B Vision
11B · 9GB · Llama 3.2 Community
Multimodal Llama: screenshots, charts and photos understood locally.
DeepSeek-R1 Distill 7B
7B · 5.6GB · MIT
Chain-of-thought reasoning distilled into a 12GB-friendly body. MIT licensed.
DeepSeek-R1 Distill 14B
14B · 10GB · MIT
The reasoning sweet spot for 16GB cards — math olympiad energy.
DeepSeek-R1 Distill 32B ★
32B · 20GB · MIT
24GB rigs' reasoning monster: shows its work, solves hard problems offline.
DeepSeek-V3 (0324)
671B · 380GB · MIT
Frontier-class open weights. Listed for completeness — you'll want a server (or Vast.ai).
DeepSeek-Coder-V2 Lite
16B · 10GB · DeepSeek License
MoE coder (2.4B active): fast fill-in-the-middle for 16GB cards.
Gemma 3 1B
1B · 1.5GB · Gemma Terms
Google's tiny entry — polished tone for its size class.
Gemma 3 4B
4B · 3.6GB · Gemma Terms
Multimodal (image-aware) at laptop-CPU-friendly size.
Gemma 3 12B ★
12B · 9GB · Gemma Terms
Vision + 128K context + great writing. A 12GB-card all-rounder.
Gemma 3 27B
27B · 17.5GB · Gemma Terms
Gemma's ceiling: single-GPU multimodal with serious writing quality.
Mistral 7B v0.3
7B · 5.5GB · Apache-2.0
The original local darling. Apache-2.0, everywhere, still useful.
Mistral Nemo 12B
12B · 8.5GB · Apache-2.0
12GB cards' multilingual workhorse with 128K context, Apache-2.0.
Mistral Small 3.1 24B
24B · 15.5GB · Apache-2.0
Apache-2.0 multimodal 24B — the license-clean 24GB flagship alternative.
Codestral 22B
22B · 14.5GB · MNPL (non-production)
Specialized coder with fill-in-the-middle; check the non-production license.
Phi-4 14B ★
14B · 9.8GB · MIT
Microsoft's textbook-trained 14B — punches far above its weight in math and logic.
Phi-4-mini 3.8B
3.8B · 3.1GB · MIT
Small, MIT-licensed, long context. A laptop staple.
QwQ 32B
32B · 20GB · Apache-2.0
Qwen's reasoning-only model — long deliberation, Apache-2.0.
Qwen2.5-VL 7B
7B · 6.5GB · Apache-2.0
Best-in-class open vision-language at 8B scale: OCR, UI understanding, video.
MiniCPM-V 2.6
8B · 6.3GB · Apache-2.0
Efficient GPT-4V-style vision that runs on phones and laptops.
IBM Granite 3.3 8B
8B · 6GB · Apache-2.0
Enterprise-minded Apache-2.0 model with strong RAG manners.
BGE-M3 ★
0.6B · 1.5GB · MIT
Multilingual embedding workhorse — the R in your local RAG.
Nomic Embed Text v1.5
0.14B · 0.8GB · Apache-2.0
Featherweight embedding model with Matryoshka dimensions.
Stable Diffusion 1.5
1B · 4GB · OpenRAIL
The ancient one that still runs on 4GB cards — endless fine-tunes and LoRAs.
SDXL Base 1.0
3.5B · 7GB · OpenRAIL++
The 12GB-card image standard with a massive fine-tune ecosystem.
FLUX.1 [schnell] ★
12B · 12GB · Apache-2.0
4-step Apache-2.0 image generation with stunning text rendering.
Whisper Large v3 ★
1.5B · 3.5GB · MIT
Speech-to-text in 99 languages, MIT, runs via whisper.cpp on CPU.
Kokoro 82M
0.08B · 1GB · Apache-2.0
Tiny TTS with shockingly natural voices — runs anywhere, even CPU-only.