OLMo 2 32B
100 trust score
Earlier fully open model. Brilliant and honest, but forgets what you said a few pages back.
Guardrails: middle What this means
Model ID, for any AI harness
allenai/OLMo-2-0325-32B-InstructTop 15 models, highest trust first
Olmo 3 7B
100 trust score
Fully open chat model. Data and code are public. Fits on the laptop you already own, if it has 8 GB.
Guardrails: middle What this means
Model ID, for any AI harness
allenai/Olmo-3-7B-InstructOlmo 3.1 32B
100 trust score
Largest fully open model here. Data and code are public. Nothing up its sleeve.
Guardrails: middle What this means
Model ID, for any AI harness
allenai/Olmo-3.1-32B-InstructSmolLM3 3B
100 trust score
Small, fully open, training data mix is public. Also has the cutest name on this list.
Guardrails: not sure yet What this means
Model ID, for any AI harness
HuggingFaceTB/SmolLM3-3BNew
Command A+ 05-2026
84 trust score
Big Cohere model for business agents, many languages, and images. Needs server GPUs.
Guardrails: not sure yet What this means
Model ID, for any AI harness
CohereLabs/command-a-plus-05-2026-bf16DeepSeek R1 0528 8B
84 trust score
A newer small thinker. MIT license, so legal can relax.
Guardrails: not sure yet What this means
Model ID, for any AI harness
deepseek-ai/DeepSeek-R1-0528-Qwen3-8BDeepSeek R1 Distill 14B
84 trust score
Thinks out loud before it answers. Get comfy.
Guardrails: not sure yet What this means
Model ID, for any AI harness
deepseek-ai/DeepSeek-R1-Distill-Qwen-14BDeepSeek R1 Distill 7B
84 trust score
Shows its work, like your math teacher always wanted. Fits in 8 GB.
Guardrails: not sure yet What this means
Model ID, for any AI harness
deepseek-ai/DeepSeek-R1-Distill-Qwen-7BNew
DeepSeek V4 Flash 0731
84 trust score
Huge model for coding agents and very long documents. Not a home-computer model.
Guardrails: not sure yet What this means
Model ID, for any AI harness
deepseek-ai/DeepSeek-V4-Flash-0731New
DeepSeek V4 Pro 0813
84 trust score
DeepSeek's largest model, for hard coding and agent jobs. It has 1.6 trillion parameters, so it needs a data center.
Guardrails: not sure yet What this means
Model ID, for any AI harness
deepseek-ai/DeepSeek-V4-Pro-0813New
DeepSeek V4.1 Flash
84 trust score
Huge multimodal MoE for agents and long context. Not a home-computer model.
Guardrails: not sure yet What this means
Model ID, for any AI harness
deepseek-ai/DeepSeek-V4.1-FlashDevstral Small
84 trust score
Fixes bugs across many files. Your code reviewer will still have notes.
Guardrails: not sure yet What this means
Model ID, for any AI harness
mistralai/Devstral-Small-2505Gemma 4 12B
84 trust score
Google's open model with an Apache license. Fits a 16 GB machine.
Guardrails: middle What this means
Model ID, for any AI harness
google/gemma-4-12B-itNew
Gemma 4 26B-A4B
84 trust score
Fast Gemma 4 that reads images. Only about 4 billion of its 26 billion parameters work at once.
Guardrails: middle What this means
Model ID, for any AI harness
google/gemma-4-26B-A4B-itNew
Gemma 4 31B
84 trust score
Google's largest Gemma 4. Strong reasoning, coding, and image reading for a 32 GB machine.
Guardrails: middle What this means
Model ID, for any AI harness
google/gemma-4-31B-it// The directory
Compare before you download.
Best lists
Best for low-RAM laptops
Models that fit in 8 GB. No upgrade, no tears.
Best for Mac
Picks for 16, 32, and 64 GB Macs. Yes, even the base model you swore was "enough."
Best for business use
Standard open licenses with no size caps. Send this one to legal.
Best fully open
Weights, code, and training data, all public. The overshare club.
Shopping for a computer?
Our buying guide starts at about $900 and goes up to the big 128 GB machines. Macs, gaming PCs, mini PCs, and NVIDIA's little AI box. Plus a quick test for the classic question: can my kid's gaming PC do this?
Starter: Mac mini
16 GB to start. Runs the small and mid models.
The gaming PC
A 16 GB NVIDIA card. Homework machine by day, AI box by night.
The big boxes
Mac Studio, DGX Spark, Ryzen AI Max. For the 120B crowd.
Plain answers to common questions
- Is Llama open source? (Short version: not by the official definition.)
- What is the best local AI for 16 GB of RAM?
- Can my gaming PC run AI?
- What do AI model licenses mean? And the glossary for the nerd words.
- New to this? Start with our setup guides.
How we check
We read each model's page and license on the maker's own site or Hugging Face account. Each fact has a source link and a check date. If we could not confirm a fact, we say "not yet checked." We do not guess. Guessing is for nerd trivia night.
What's new
- New model Mistral Large 4: Added Mistral AI's Mistral Large 4. A huge model for coding, agents, and reading images, with about 1 trillion parameters. Mistral Ecosystem Open-Weight License; business use is fine unless you sell model access as a service. No sign-up needed. Not a home-computer model.
- New model LFM3 1.7B: Added Liquid AI's LFM3 1.7B. A tiny, fast model for agents and tool use on a phone or laptop, about 1 GB at 4-bit. LFM Open License v1.0; business use only under 10 million US dollars a year. No sign-up needed.
- New model LFM3 VL 2B: Added Liquid AI's LFM3 VL 2B. A small model that reads images and video on your own device. LFM Open License v1.0; business use only under 10 million US dollars a year. No sign-up needed.
- New model MiniMax M3.1 Flash: Added MiniMax M3.1 Flash. A huge coding and agent model that reads images and video, with 23 billion of its 428 billion parameters active at once. MiniMax Community License; business use allowed with limits. No sign-up needed.
- New model Nemotron 3.5 Bolt E4B: Added NVIDIA's Nemotron 3.5 Bolt E4B. A small, fast helper for voice and device assistants that use tools. Built on Gemma 4 E4B. OpenMDW 1.1; business use allowed. No sign-up needed.
Every Friday: new models, score changes, and the license gotchas before they bite. One email, three minutes.
Opens our sign-up page on Beehiiv, our email provider. Confirm your email there to join.