Find the Best Local AI Model for Your Device
ModelFit is a free tool that tells you which local AI models run on your hardware. Tell us your hardware and get the single best local LLM it can actually run, sized to your exact memory. Honest speed and quality estimates for Apple Silicon, iPhone, NVIDIA, and AMD.
How fast is your machine, really?
One command shows your real tok/s, then a card to post on X and your rank on the leaderboard. It asks before it sends anything.
No Node? brew tap modelfit-io/tap, brew trust modelfit-io/tap, brew install modelfit, then modelfit bench. The first run downloads 3 to 17 GB.
- Models tracked
- 200
- Run locally
- 156
- Devices and GPUs
- 71
~ $ npx @modelfit/cli@latest bench
run 1/3: 132.6 tok/s
run 2/3: 155.3 tok/s
run 3/3: 150.0 tok/s
150.0 tok/s on Qwen3.8 27B, #1 of 12
Your card, ready to post on X:
A real run from the leaderboard, Nov 10.
Choose your device
We match models to your exact memory and architecture.
Pick your model in one line, measure your machine in another
ModelFit's free, zero-dependency CLI detects your exact hardware, picks the single best local model it can run, and hands you the one-line Ollama command to install it. The second command times a model that fits your machine with Ollama and asks before adding your speed to the public leaderboard below. The dataset behind it is open (CC BY 4.0), and AI agents can query it live over MCP.
- 1 · Pick the best model$npx @modelfit/cli@latest
- 2 · Measure your real speed$npx @modelfit/cli@latest bench
Where does your machine land?
Full leaderboard →Every other speed on ModelFit is an estimate. These are real runs: readers time Qwen3.8 27B with the same prompts on their own Mac or GPU, so the bars compare. 15 runs from 12 machines so far.
Not on the board yet: M5, RTX 4090, RTX 3090, M3, RTX 4060, M2, RTX 4070, M3 Pro, RTX 5080, M3 Max. Own one? The second command above adds it. 4 more machines are on the full leaderboard.
Median of each machine's runs, generation speed as reported by Ollama. Needs Node and Ollama. The CLI asks before it sends anything: machine type, memory, Ollama version and timings (no name, no email, no files). New runs reach this chart at the next site update and the full leaderboard right away.
Latest from the blog
View all →
Nov 14, 2026llama.cpp 1.0 on a Mac: CLI Keeps Working, Ripple 65B Needs 64GB
llama.cpp 1.0.0 shipped on November 13, 2026, and for most Mac users nothing breaks. The release removes 35 deprecated functions from the llama.h C API, so apps and language bindings that still call t...
- Nov 13, 2026
Gemini 4 Hydrogen Is Cloud Only: No Weights, No Local Run
Google DeepMind introduced Gemini 4 Hydrogen on 12 November 2026 (Google DeepMind, 2026). It is the fast member of the Gemini 4 family, built for everyday coding and agentic workflows, next to the fro...
- Nov 13, 2026
Can a 512GB Mac Run Inkling 2 Small? 500GB Card Minimum
Short answer: the weights are small enough for a 512GB Mac Studio on paper, and no Mac engine is documented to load them. Thinking Machines opened the Inkling 2 Small weights on 12 November under Apac...
- Nov 12, 2026
ERNIE 5.5 Flash Preview: API Today, Open Weights Before Nov 25
Baidu released ERNIE 5.5 Flash Preview on 11 November 2026, the first post-trained version of the ERNIE 5.5 Flash Base checkpoint it published on 5 November. It has 280B total parameters, 14B active, ...
How much LLM fits your memory?
All stats →A local LLM needs roughly 0.6 GB per billion parameters at Q4. ModelFit sizes picks to ~70% of your unified memory on machines up to 32GB, scaling to ~85% at 128GB+, leaving room for the OS, context, and KV-cache. Here is what each tier comfortably runs today. The same math applies to GPU VRAM: for dedicated cards, see the per-card GPU guides.
| Unified memory | Runs up to (dense) | Top local pick | Models that fit |
|---|---|---|---|
| 8 GB | ~9B params | Qwen3.5 4B Instruct | 36 of 156 |
| 12 GB | ~12B params / 12.1B MoE | LFM2.5 8B-A1B | 56 of 156 |
| 16 GB | ~14B params | Qwen3.5 9B Instruct | 67 of 156 |
| 24 GB | ~29.3B params | LFM2 24B-A2B Instruct | 82 of 156 |
| 32 GB | ~33.1B params / 35.95B MoE | Qwen3.8 27B | 106 of 156 |
| 36 GB | ~33.1B params / 39.8B MoE | Qwen3.6 35B-A3B | 111 of 156 |
| 48 GB | ~33.1B params / 46.7B MoE | Qwen3.6 35B-A3B | 121 of 156 |
| 64 GB | ~70B params / 78.1B MoE | Nemotron 3.5 Lightning 30B-A3B | 130 of 156 |
| 72 GB | ~70B params / 80B MoE | Kolibri 1 (IQ4_XS) | 132 of 156 |
| 96 GB | ~123B params / 125.09B MoE | Kolibri 1 | 142 of 156 |
| 128 GB | ~123B params / 125.09B MoE | Qwen3.5 122B-A10B Instruct | 148 of 156 |
| 192 GB | ~123B params / 235B MoE | Qwen3.8-Flash-Next | 150 of 156 |
| 256 GB | ~123B params / 428B MoE | Qwen3.8-Flash-Next | 151 of 156 |
| 512 GB | ~123B params / 671B MoE | MiniMax M3.1 Flash | 156 of 156 |
Derived from ModelFit's own catalog of 200 models · dataset updated 2026-11-12 · full dataset (CC BY 4.0)
Run your own number: VRAM calculator, per-GPU answers, VRAM per model size, RAM vs VRAM or the full memory calculator.
Local vs cloud: how close are we?
Full benchmark →Open-weight models you can run at home keep narrowing the gap with the best closed APIs on coding. Here is where the line sits today on SWE-Bench Verified.
SWE-Bench Verified · third-party scores, each raw-confirmed against the model's primary source. Qwen3.8-27B runs locally on a 24GB Mac.
Browse the full index
All devices →Local LLM questions, answered
How much memory do I need to run a local LLM?
A local model needs roughly 0.6 GB of memory per billion parameters at Q4 quantization for the weights, closer to 0.8 GB per billion under 8B once runtime overhead is counted. ModelFit sizes its picks to about 70% of your device's unified memory (scaling to ~85% on 128GB+ machines), so an 8GB machine comfortably runs dense models up to ~9B and a 16GB machine up to ~14B parameters. Mixture-of-experts models fit larger total parameter counts because only a fraction is active per token.
What is the best local LLM for a 16GB Mac?
On a 16GB Mac, ModelFit's top pick is Qwen3.5 9B Instruct (Q4_K_M): it loads in about 7GB and ranks first on the engine's balance of quality, speed and memory headroom, the same pick as the calculator. 67 of ModelFit's 156 local models fit a 16GB device.
Can I run these models completely offline?
Yes. Every local model ModelFit recommends runs fully offline on your own hardware, through Ollama or, for builds without an Ollama tag, llama.cpp, LM Studio or MLX: no API keys, no subscription, and no data leaves your machine.
How many AI models does ModelFit cover?
ModelFit tracks 200 AI models across 40 families; 156 run locally on Apple Silicon, NVIDIA, or AMD hardware, 125 of them with a registry-verified Ollama tag.
How accurate are the speed estimates?
Tokens-per-second figures are ModelFit estimates derived from your chip or GPU's memory bandwidth and the model's size, not measured benchmarks. Real-world speed varies with quantization, context length, and thermal state.