MODELS BY RAM

Best local AI models for 32 GB RAM

At 32 GB of RAM, you can run virtually every local AI model — including 7B+ parameter reasoning, code, and vision models at high quantization quality. This is the enthusiast tier.

71 models run · 4 too heavy · Ultra tier
BROWSE BY TIER
Low tier
4 GB
Mid tier
8 GB
High tier
16 GB
Ultra tier
32 GB
FILTER BY TASK
Chat 32 models
Ternary Bonsai 8B
Ternary Bonsai 8B 1.6 GB
PrismML's 1.58-bit ternary model — top intelligence at 8B in ~1.6 GB
Q2 ~125 tok/s
Tight fit → Qwen 3.5 2B
Qwen 3.5 2B 1.3 GB
Compact multimodal chat — native vision, 262K context
FP16 ~48 tok/s
Runs great → Llama 3.2 1B
Llama 3.2 1B 791 MB
Meta's solid all-rounder
FP16 ~74 tok/s
Runs great → Bonsai 8B (1-bit)
Bonsai 8B (1-bit) 1.16 GB
PrismML's native 1-bit model — 8B params in 1.16 GB, needs forked llama.cpp
Q1 ~172 tok/s
Tight fit → Gemma 4 E2B
Gemma 4 E2B 1.5 GB
Google's 2026 on-device model — 2.3B active / 5.1B total, multimodal vision and audio
FP16 ~44 tok/s
Runs great → Qwen 3.5 0.8B
Qwen 3.5 0.8B 560 MB
Alibaba's tiny multimodal model — vision-capable, 262K context, runs on phones
FP16 ~111 tok/s
Runs great → Gemma 3 12B
Gemma 3 12B 7.3 GB
Google's largest open model — near-frontier quality
Q8 ~15 tok/s
Runs great → Mistral Nemo 12B
Mistral Nemo 12B 7.1 GB
Mistral's 12B — Tekken tokenizer, 128K context
Q8 ~16 tok/s
Runs great → Gemma 3 1B
Gemma 3 1B 786 MB
Google's compact model — great reasoning for its size
FP16 ~91 tok/s
Runs great → LFM2.5 1.2B
LFM2.5 1.2B 731 MB
Liquid AI's hybrid model — blazing fast CPU inference
FP16 ~85 tok/s
Runs great → Qwen 2.5 0.5B
Qwen 2.5 0.5B 477 MB
Lightweight chat — surprisingly capable
FP16 ~167 tok/s
Runs great → Llama 3.2 3B
Llama 3.2 3B 2.0 GB
Meta's 3B — best small model for many tasks
FP16 ~30 tok/s
Runs great → SmolLM2 1.7B
SmolLM2 1.7B 1.0 GB
Largest SmolLM2 — punches above its weight
FP16 ~59 tok/s
Runs great → Gemma 3 4B
Gemma 3 4B 2.8 GB
Google's 4B — strong reasoning and instruction following
FP16 ~22 tok/s
Runs great → Qwen 3.5 4B
Qwen 3.5 4B 2.5 GB
Multimodal sweet spot — vision, strong reasoning, 262K context
FP16 ~24 tok/s
Runs great → Ternary Bonsai 4B
Ternary Bonsai 4B 860 MB
PrismML's 1.58-bit ternary model — 4B-class intelligence in ~0.86 GB, ~9x smaller than fp16
Q2 ~200 tok/s
Tight fit → Gemma 4 E4B
Gemma 4 E4B 3.0 GB
Google's most capable edge model — 4B effective, multimodal, 128K context
FP16 ~22 tok/s
Runs great → SmolLM2 360M
SmolLM2 360M 267 MB
Ultra-small language model
FP16 ~200 tok/s
Runs great → Qwen 2.5 3B
Qwen 2.5 3B 2.0 GB
Sweet-spot model — great quality-to-size ratio
FP16 ~31 tok/s
Runs great → SmolLM2 135M
SmolLM2 135M 99 MB
Ultra-tiny language model — runs anywhere
FP16 ~200 tok/s
Runs great → StableLM 2 1.6B
StableLM 2 1.6B 1.0 GB
Stability AI's efficient small model
FP16 ~61 tok/s
Runs great → Qwen 2.5 7B
Qwen 2.5 7B 4.4 GB
Top-tier 7B chat — rivals much larger models
FP16 ~13 tok/s
Runs well → Llama 3.1 8B
Llama 3.1 8B 4.9 GB
Meta's workhorse 8B — excellent all-around
FP16 ~12 tok/s
Runs well → TinyLlama 1.1B
TinyLlama 1.1B 780 MB
Popular tiny model trained on 3T tokens
FP16 ~91 tok/s
Runs great → Ternary Bonsai 1.7B
Ternary Bonsai 1.7B 380 MB
PrismML's 1.58-bit ternary model — 1.7B params in under 0.4 GB, runs anywhere
Q2 ~200 tok/s
Tight fit → Gemma 3n E4B
Gemma 3n E4B 4.5 GB
Google's most capable on-device model — 4B effective, multimodal
FP16 ~15 tok/s
Runs great → Gemma 3n E2B
Gemma 3n E2B 3.0 GB
Google's on-device multimodal — 2B effective params with vision and audio
FP16 ~22 tok/s
Runs great → Qwen 3.5 9B
Qwen 3.5 9B 5.5 GB
Flagship small Qwen — native multimodal, 262K context, rivals far larger models
FP16 ~11 tok/s
Tight fit → Gemma 4 31B
Gemma 4 31B 18 GB
Google's dense flagship open model — near server-grade quality, 256K context
Q5 ~10 tok/s
Tight fit → LFM2.5 8B A1B
LFM2.5 8B A1B 4.7 GB
Liquid AI's on-device MoE — 8.3B total, 1.5B active, reasoning + tool calling at 3–4B quality
FP16 ~12 tok/s
Runs well → Gemma 4 26B A4B
Gemma 4 26B A4B 16 GB
Google's sparse MoE — 26B total, 3.8B active per token, 256K context
Q6 ~10 tok/s
Tight fit → Qwen 3.5 35B A3B
Qwen 3.5 35B A3B 20 GB
Alibaba's sparse MoE — 35B total, 3B active per token, multimodal, 262K context
Q4 ~10 tok/s
Tight fit →
Reasoning 7 models
🧠
DeepSeek R1 Distill 1.5B 1.1 GB
Distilled reasoning — chain-of-thought in a tiny package
FP16 ~61 tok/s
Runs great → Phi-4 Mini
Phi-4 Mini 8.7 GB
Microsoft's reasoning model — exceptional for its size
FP16 ~26 tok/s
Runs great → 🧠
DeepSeek R1 Distill 7B 4.7 GB
Distilled from DeepSeek R1 — strong step-by-step reasoning
FP16 ~13 tok/s
Runs well → 🧠
DeepSeek R1 Distill 8B 4.9 GB
Llama-based R1 distill — best open reasoning at 8B
FP16 ~12 tok/s
Runs well → 🧠
Qwen3 8B 5.2 GB
Alibaba's latest — thinking mode with strong reasoning
FP16 ~12 tok/s
Runs well → Phi-4 Medium
Phi-4 Medium 8.0 GB
Microsoft's 14B reasoning model — frontier-class performance
Q8 ~14 tok/s
Runs well → Mistral 7B
Mistral 7B 4.3 GB
High-quality reasoning and analysis
FP16 ~14 tok/s
Runs well →
Code 7 models
Qwen 2.5 Coder 1.5B
Qwen 2.5 Coder 1.5B 1.1 GB
Best small code model — great for autocomplete
FP16 ~63 tok/s
Runs great → Qwen 2.5 Coder 0.5B
Qwen 2.5 Coder 0.5B 477 MB
Tiny code completion model — autocomplete on any device
FP16 ~167 tok/s
Runs great → Qwen 2.5 Coder 3B
Qwen 2.5 Coder 3B 2.0 GB
Strong code generation and editing at 3B
FP16 ~31 tok/s
Runs great → Qwen 2.5 Coder 7B
Qwen 2.5 Coder 7B 4.4 GB
Best open 7B code model — rivals GPT-4 on coding benchmarks
FP16 ~13 tok/s
Runs well → 👨‍💻
StarCoder2 3B 1.9 GB
BigCode's multilingual code model — 600+ languages
FP16 ~32 tok/s
Runs great → 👨‍💻
DeepSeek Coder 6.7B 3.8 GB
DeepSeek's code model — strong at generation and debugging
FP16 ~15 tok/s
Runs great → 👨‍💻
Laguna XS.2 19 GB
Poolside's open agentic coding MoE — 33B total, 3B active, runs locally on 36GB Macs
Q5 ~9 tok/s
Tight fit →
Vision 4 models
👁️
Moondream 2B 1.3 GB
Small but capable vision model — image Q&A and captioning
FP16 ~54 tok/s
Runs great → 👁️
SmolVLM 500M 490 MB
Tiny vision-language model — describe images on any device
FP16 ~182 tok/s
Runs great → Qwen 2.5 VL 7B
Qwen 2.5 VL 7B 4.6 GB
State-of-the-art vision-language — image, video, document understanding
FP16 ~12 tok/s
Runs well → 👁️
LLaVA 1.6 7B 4.5 GB
Leading open vision model — image understanding and reasoning
FP16 ~14 tok/s
Runs well →
Embedding 4 models
🔢
Nomic Embed Text 137 MB
Open-source embedding with 8K context — long document search
FP16 ~200 tok/s
Runs great → 🔢
GTE Large 335 MB
High-quality embeddings — top of MTEB benchmark at its size
FP16 ~200 tok/s
Runs great → 🔢
all-MiniLM-L6-v2 23 MB
Fast sentence embeddings — ideal for semantic search
FP16 ~200 tok/s
Runs great → 🔢
BGE Small 34 MB
Compact BAAI embedding — great for RAG pipelines
FP16 ~200 tok/s
Runs great →
Image 5 models
🎨
Stable Diffusion Turbo 2.2 GB
1-step image generation — instant results
FP16 ~95 tok/s
Runs great → 🎨
Stable Diffusion 3.5 Medium 4.6 GB
Latest SD architecture — excellent image quality
FP16 ~43 tok/s
Runs great → 🎨
FLUX.1 Schnell 12.1 GB
Black Forest Labs' fast model — stunning quality in 4 steps
Q8 ~16 tok/s
Runs great → Bonsai Image 4B
Bonsai Image 4B 1.8 GB
PrismML's 1.58-bit ternary diffusion — first 4B image model to run on iPhone, ~3GB in-browser
Q2 ~111 tok/s
Tight fit → 🎨
SDXL Turbo 6.5 GB
High-res 1-step generation — 1024×1024 in seconds
FP16 ~31 tok/s
Runs great →
Voice 6 models
🗣️
Dia 1.6B 1.6 GB
Nari Labs dialogue TTS — multi-speaker with emotion
FP16 ~61 tok/s
Runs great → 🗣️
NeuTTS Air 450 MB
Neuphonic's on-device TTS — 748M, instant voice cloning, runs on CPU via llama.cpp
FP16 ~133 tok/s
Runs great → 🗣️
Kokoro 82M 183 MB
24 natural voices — instant TTS
FP16 ~200 tok/s
Runs great → 🗣️
OuteTTS 0.3 500M 500 MB
Voice cloning and natural TTS — zero-shot voice synthesis
FP16 ~200 tok/s
Runs great → 🗣️
KittenTTS Mini 80 MB
KittenML's expressive 80M TTS — 8 voices, quantization-aware, runs on any CPU
FP16 ~200 tok/s
Runs great → 🗣️
KittenTTS Nano 19 MB
Tiny, fast voice synthesis
FP16 ~200 tok/s
Runs great →
Transcription 6 models
Whisper Large V3
Whisper Large V3 3.1 GB
Best open transcription — near-human accuracy
FP16 ~65 tok/s
Runs great → Whisper Medium
Whisper Medium 1.5 GB
Strong multilingual transcription
FP16 ~133 tok/s
Runs great → 🎙️
Distil-Whisper Large V3 1.5 GB
6x faster than Whisper Large — nearly same accuracy
FP16 ~133 tok/s
Runs great → Wav2Vec2 Base
Wav2Vec2 Base 231 MB
High-accuracy English ASR
FP16 ~200 tok/s
Runs great → Whisper Small
Whisper Small 488 MB
Good accuracy-speed balance — 99 languages
FP16 ~200 tok/s
Runs great → Whisper Tiny
Whisper Tiny 89 MB
Real-time speech recognition — runs on anything
FP16 ~200 tok/s
Runs great →
Too heavy 4 models
DeepSeek V4 Flash
DeepSeek V4 Flash 156 GB
DeepSeek's V4 Flash MoE — 284B total, 13B active, 1M context (server-grade hardware)
needs 95.0 GB
Too heavy → Kimi K2.6
Kimi K2.6 594 GB
Moonshot's 1T MoE — 32B active, frontier agentic coding; needs a multi-GPU workstation
needs 300.0 GB
Too heavy → Llama 4 Scout
Llama 4 Scout 63 GB
Meta's MoE model — 109B total params, 17B active per token
needs 42.0 GB
Too heavy → Qwen3 Coder Next
Qwen3 Coder Next 46 GB
Alibaba's local coding agent MoE — 80B total, 3B active, agentic long-horizon coding, 256K context
needs 28.0 GB
Too heavy →