Local models - verified
Open-weight models you can run locally
63 open-weight models from 19 vendors: how big each download is, which quantizations exist, what the license lets you do and the exact command to run it, copied from official pages.
Picking the software to run them? Compare 12 local LLM runtimes: Ollama, LM Studio, llama.cpp, vLLM and more, by platform, API and features.
How much memory do you have?
Pick your GPU memory or unified memory. Each tier lists the models whose smallest listed download uses at most 85% of it, leaving room for the context and the OS.
- Models
- 63
- Vendors
- 19
- Commercial: yes
- 48
- Ollama tags
- 36
Every model
Smallest download is the smallest listed file: a quantization, or the official weights when nothing smaller is listed. Sorted smallest first; click a column to re-sort.
63 models
| Commercial | Ollama | |||||
|---|---|---|---|---|---|---|
Qwen3.5-0.8BQwen/Qwen3.5-0.8B | Alibaba Cloud (Qwen) | 0.9B | 0.5 GBQ4_K_M GGUF | 262K | Commercial use: Yes | qwen3.5:0.8b |
Qwen3.5-2BQwen/Qwen3.5-2B | Alibaba Cloud (Qwen) | 2.3B | 1.3 GBQ4_K_M GGUF | 262K | Commercial use: Yes | qwen3.5:2b |
Liquid AI LFM2.5-2.6BLiquidAI/LFM2.5-2.6B | Liquid AI | 2.7B | 1.6 GB4bit MLX | 131K | Commercial use: Conditions | -No Ollama tag |
Ministral 3 3B Instruct 2512mistralai/Ministral-3-3B-Instruct-2512 | Mistral AI | 3.8B | 2.1 GBQ4_K_M GGUF | 262K | Commercial use: Yes | ministral-3:3b |
Phi-4-mini-instructmicrosoft/Phi-4-mini-instruct | Microsoft | 3.8B | 2.2 GB4bit MLX | 131K | Commercial use: Yes | phi4-mini:3.8b |
Phi-4-mini-reasoningmicrosoft/Phi-4-mini-reasoning | Microsoft | 3.8B | 2.2 GB4bit MLX | 131K | Commercial use: Yes | phi4-mini-reasoning:3.8b |
Qwen3.5-4BQwen/Qwen3.5-4B | Alibaba Cloud (Qwen) | 4.7B | 2.7 GBQ4_K_M GGUF | 262K | Commercial use: Yes | qwen3.5:4b |
Gemma 4 E2Bgoogle/gemma-4-E2B-it | 5.1B | 3.1 GBQ4_K_M GGUF | 131K | Commercial use: Yes | gemma4:e2b | |
Cohere Command R7B (12-2024)CohereLabs/c4ai-command-r7b-12-2024 | Cohere | 8B | 4.5 GB4bit MLX | 131K | Commercial use: No | command-r7b:7b |
IBM Granite 4.2 8Bibm-granite/granite-4.2-8b | IBM | 8.8B | 4.9 GB4bit MLX | 131K | Commercial use: Yes | granite4.2:8b |
Gemma 4 E4Bgoogle/gemma-4-E4B-it | 8B | 5.0 GBQ4_K_M GGUF | 131K | Commercial use: Yes | gemma4:e4b | |
Ministral 3 8B Instruct 2512mistralai/Ministral-3-8B-Instruct-2512 | Mistral AI | 8.9B | 5.2 GBQ4_K_M GGUF | 262K | Commercial use: Yes | ministral-3:8b |
Qwen3.5-9BQwen/Qwen3.5-9B | Alibaba Cloud (Qwen) | 9.7B | 5.7 GBQ4_K_M GGUF | 262K | Commercial use: Yes | qwen3.5:9b |
Gemma 4 12Bgoogle/gemma-4-12B-it | 12B | 6.7 GB4bit MLX | 262K | Commercial use: Yes | gemma4:12b | |
Ministral 3 14B Instruct 2512mistralai/Ministral-3-14B-Instruct-2512 | Mistral AI | 13.9B | 8.2 GBQ4_K_M GGUF | 262K | Commercial use: Yes | ministral-3:14b |
Phi-4microsoft/phi-4 | Microsoft | 14.7B | 8.2 GB4bit MLX | 16K | Commercial use: Yes | phi4:14b |
Phi-4-reasoning-plusmicrosoft/Phi-4-reasoning-plus | Microsoft | 14.7B | 9.1 GBQ4_K_M GGUF | 33K | Commercial use: Yes | phi4-reasoning:plus |
gpt-oss-20bopenai/gpt-oss-20b | OpenAI | 20.9B3.6B active | 11.2 GBMXFP4-Q4 MLX | 131K | Commercial use: Yes | gpt-oss:20b |
Liquid AI LFM2.5-24B-A2BLiquidAI/LFM2.5-24B-A2B | Liquid AI | 24B2.5B active | 13.6 GB4bit MLX | 131K | Commercial use: Conditions | lfm2.5:24b |
Devstral Small 2 24B Instruct 2512mistralai/Devstral-Small-2-24B-Instruct-2512 | Mistral AI | 24B | 14.3 GBQ4_K_M GGUF | 262K | Commercial use: Yes | devstral-small-2:24b |
Gemma 4 26B A4Bgoogle/gemma-4-26B-A4B-it | 25.8B3.8B active | 14.4 GBQ4_0 (QAT) GGUF | 262K | Commercial use: Yes | gemma4:26b | |
Ai2 Olmo Hybrid Instruct DPO 7Ballenai/Olmo-Hybrid-Instruct-DPO-7B | Ai2 | 7.4B | 14.9 GBOfficial BF16 | 33K | Commercial use: Yes | -No Ollama tag |
Qwen3.8-27BQwen/Qwen3.8-27B | Alibaba Cloud (Qwen) | 27.8B | 16.1 GB4bit MLX | 262K | Commercial use: Yes | qwen3.8:27b |
IBM Granite 4.2 30Bibm-granite/granite-4.2-30b | IBM | 29.3B | 16.5 GB4bit MLX | 131K | Commercial use: Yes | granite4.2:30b |
Gemma 4 31Bgoogle/gemma-4-31B-it | 31.3B | 17.7 GBQ4_0 (QAT) GGUF | 262K | Commercial use: Yes | gemma4:31b | |
NVIDIA Nemotron 3.5 Lightning 30B-A3Bnvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16 | NVIDIA | 31.6B3B active | 17.8 GB4bit MLX | 1.05M | Commercial use: Yes | nemotron-3.5-lightning:30b |
Ai2 Olmo 3.1 32B Thinkallenai/Olmo-3.1-32B-Think | Ai2 | 32.2B | 18.1 GB4bit MLX | 66K | Commercial use: Yes | olmo-3.1:32b-think |
Cohere Aya Expanse 32BCohereLabs/aya-expanse-32b | Cohere | 32.3B | 19.8 GBQ4_K_M GGUF | 131K | Commercial use: No | aya-expanse:32b |
Qwen3.6-35B-A3BQwen/Qwen3.6-35B-A3B | Alibaba Cloud (Qwen) | 36B3B active | 20.4 GB4bit MLX | 262K | Commercial use: Yes | qwen3.6:35b |
NVIDIA Nemotron 3.5 Bolt 74B-A9Bnvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-BF16 | NVIDIA | 74.9B9B active | 22.0 GB24GB (native 2-bit) GGUF | 262K | Commercial use: Yes | nemotron-3.5-bolt:74b-a9b |
Phi-4-reasoning-vision-15Bmicrosoft/Phi-4-reasoning-vision-15B | Microsoft | 15.1B | 30.2 GBOfficial BF16 | 16K | Commercial use: Yes | -No Ollama tag |
AI21 Jamba2 Miniai21labs/AI21-Jamba2-Mini | AI21 Labs | 51.6B12B active | 31.2 GBQ4_K_M GGUF | 262K | Commercial use: Yes | -No Ollama tag |
Qwen3-Coder-NextQwen/Qwen3-Coder-Next | Alibaba Cloud (Qwen) | 79.7B3B active | 44.8 GB4bit MLX | 262K | Commercial use: Yes | qwen3-coder-next |
GLM-4.5-Airzai-org/GLM-4.5-Air | Z.ai | 110.5B12B active | 60.1 GB4bit MLX | 131K | Commercial use: Yes | -No Ollama tag |
Llama 4 Scout 17B-16E Instructmeta-llama/Llama-4-Scout-17B-16E-Instruct | Meta | 108.6B17B active | 61.1 GB4bit MLX | 10M | Commercial use: Conditions | llama4:scout |
GLM-4.6Vzai-org/GLM-4.6V | Z.ai | 107.7B | 61.9 GB4bit MLX | 128K | Commercial use: Yes | -No Ollama tag |
gpt-oss-120bopenai/gpt-oss-120b | OpenAI | 116.8B5.1B active | 62.3 GBMXFP4-Q4 MLX | 131K | Commercial use: Yes | gpt-oss:120b |
Cohere Command A (03-2025)CohereLabs/c4ai-command-a-03-2025 | Cohere | 111.1B | 62.5 GB4bit MLX | 262K | Commercial use: No | command-a:111b |
Mistral Small 4 119B (2603)mistralai/Mistral-Small-4-119B-2603 | Mistral AI | 119.4B6.5B active | 67.8 GB4bit MLX | 262K | Commercial use: Yes | -No Ollama tag |
Devstral 2 123B Instruct 2512mistralai/Devstral-2-123B-Instruct-2512 | Mistral AI | 125B | 70.3 GB4bit MLX | 262K | Commercial use: Conditions | devstral-2:123b |
NVIDIA Nemotron 3.5 Super VL 120B-A12Bnvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16 | NVIDIA | 124.5B12B active | 71.8 GBQ4_0 GGUF | 1.05M | Commercial use: Yes | nemotron-3.5-super:120b |
Mistral Medium 3.5 128Bmistralai/Mistral-Medium-3.5-128B | Mistral AI | 127.7B | 74.9 GBQ4_K_M GGUF | 262K | Commercial use: Conditions | mistral-medium-3.5:128b |
Step 3.7 Flashstepfun-ai/Step-3.7-Flash | StepFun | 201.4B11B active | 111.5 GBQ4_K_S GGUF | 262K | Commercial use: Yes | -No Ollama tag |
MiMo-V2.6-Flash-RLXiaomiMiMo/MiMo-V2.6-Flash-RL | Xiaomi MiMo | 310.8B15B active | 167.4 GBMXFP4 GGUF | 1.05M | Commercial use: Yes | -No Ollama tag |
Hy3tencent/Hy3 | Tencent | 298.8B21B active | 182.3 GBQ4_K_M GGUF | 262K | Commercial use: Yes | -No Ollama tag |
GLM-4.6zai-org/GLM-4.6 | Z.ai | 356.8B | 198.6 GB4bit MLX | 200K | Commercial use: Yes | -No Ollama tag |
GLM-4.7zai-org/GLM-4.7 | Z.ai | 358.3B | 198.6 GB4bit MLX | 203K | Commercial use: Yes | -No Ollama tag |
GLM-5.3-Flashzai-org/GLM-5.3-Flash | Z.ai | 321.3B18B active | 199.7 GBUD-Q4_K_XL GGUF | 1.05M | Commercial use: Yes | -No Ollama tag |
MiniMax-M3MiniMaxAI/MiniMax-M3 | MiniMax | 427B23B active | 241.5 GB4bit MLX | 1.05M | Commercial use: Conditions | -No Ollama tag |
Llama 4 Maverick 17B-128E Instructmeta-llama/Llama-4-Maverick-17B-128E-Instruct | Meta | 401.6B17B active | 242.8 GBQ4_K_M GGUF | 1M | Commercial use: Conditions | llama4:maverick |
Mistral Large 3 675B Instruct 2512mistralai/Mistral-Large-3-675B-Instruct-2512 | Mistral AI | 675B41B active | 407.0 GBQ4_K_M GGUF | 262K | Commercial use: Yes | -No Ollama tag |
GLM-5.2zai-org/GLM-5.2 | Z.ai | 753.3B | 418.3 GB4bit MLX | 1M | Commercial use: Yes | -No Ollama tag |
GLM-5.3zai-org/GLM-5.3 | Z.ai | 753.3B | 418.3 GB4bit MLX | 1.05M | Commercial use: Conditions | -No Ollama tag |
GLM-5zai-org/GLM-5 | Z.ai | 753.9B40B active | 418.6 GB4bit MLX | 203K | Commercial use: Yes | -No Ollama tag |
Hy4 previewtencent/Hy4-preview | Tencent | 780B49B active | 433.4 GB4bit MLX | 1.05M | Commercial use: Yes | -No Ollama tag |
GLM-5.1zai-org/GLM-5.1 | Z.ai | 753.9B | 464.5 GBUD-Q4_K_M GGUF | 203K | Commercial use: Yes | -No Ollama tag |
DeepSeek-V4.1-Flashdeepseek-ai/DeepSeek-V4.1-Flash | DeepSeek | 763.2B16B active | 510.3 GBOfficial Mixed FP8 (F8_E4M3), I8 and BF16 | 1M | Commercial use: Yes | -No Ollama tag |
MiMo-V2.6-Pro-RLXiaomiMiMo/MiMo-V2.6-Pro-RL | Xiaomi MiMo | 1.02T42B active | 554.2 GBMXFP4 GGUF | 1.05M | Commercial use: Yes | -No Ollama tag |
Kimi K2.6moonshotai/Kimi-K2.6 | Moonshot AI | 1.03T32B active | 583.7 GBUD-Q4_K_XL GGUF | 262K | Commercial use: Conditions | -No Ollama tag |
Kimi K2.7 Codemoonshotai/Kimi-K2.7-Code | Moonshot AI | 1.03T32B active | 583.7 GBUD-Q4_K_XL GGUF | 262K | Commercial use: Conditions | -No Ollama tag |
AI21 Jamba Large 1.7ai21labs/AI21-Jamba-Large-1.7 | AI21 Labs | 398.6B94B active | 797.1 GBOfficial BF16 | 262K | Commercial use: Conditions | -No Ollama tag |
DeepSeek-V4-Prodeepseek-ai/DeepSeek-V4-Pro | DeepSeek | 1.6T49B active | 837.0 GB4bit MLX | 1M | Commercial use: Yes | -No Ollama tag |
Kimi K3moonshotai/Kimi-K3 | Moonshot AI | 2.78T104B active | 1,508.7 GBUD-Q4_K_XL GGUF | 1.05M | Commercial use: Conditions | -No Ollama tag |
Method
How to read these numbers
Read from the Hugging Face API and file trees, model cards and ollama.com on . Repos change; check the linked source before you download hundreds of gigabytes.
- File sizes, not memory use
- Sizes are the file sizes Hugging Face reports for each repo, in decimal GB (1 GB = 10^9 bytes). Running a model needs more than the file: memory for the context (KV cache, which grows with every token you keep) and for the operating system.
- The 85% rule of thumb
- A model is listed under a memory tier when its smallest listed download is at most 85% of that memory. It is a starting point, not a guarantee: a long context or a busy machine can need more.
- Not a benchmark
- Nothing here measures quality or speed. Tables sort by size, parameters or context only. Licenses are summarized as yes, conditions or no for commercial use; read the linked license before you ship.
- Commands copied verbatim
- Every command comes from the model card, the vendor or the Ollama library page it links to. Models whose official pages publish no command show none.
FAQ
How much memory do I need to run an open model locally?
At least the size of the file you download, plus room for the context (the KV cache grows with every token you keep) and for the operating system. This site's rule of thumb: a model fits a memory tier when its smallest listed download is at most 85% of that memory, leaving the rest for the context (KV cache) and the operating system. The sizes on this page are file sizes from Hugging Face, not measured memory use.
Which open models fit in 16 GB?
By this site's 85% rule of thumb, 19 models of 63 have a listed download of 13.6 GB or less. By total parameters, largest first: Liquid AI LFM2.5-24B-A2B, gpt-oss-20b, Phi-4, Phi-4-reasoning-plus, Ministral 3 14B Instruct 2512 and 14 more. The full list with each download is at https://www.developersdigest.tech/local-models/fits/16.
Which of these models can I use commercially?
48 of 63 ship under licenses that allow commercial use outright, mostly Apache 2.0 and MIT. 12 allow it with conditions written into the license (GLM-5.3, Kimi K2.6, Kimi K2.7 Code, Kimi K3, Mistral Medium 3.5 128B, Devstral 2 123B Instruct 2512, MiniMax-M3, Llama 4 Scout 17B-16E Instruct, Llama 4 Maverick 17B-128E Instruct, Liquid AI LFM2.5-24B-A2B, Liquid AI LFM2.5-2.6B and AI21 Jamba Large 1.7), and 3 do not (Cohere Command A (03-2025), Cohere Command R7B (12-2024) and Cohere Aya Expanse 32B). Each model page links the license text; read it before you ship.
Which models can I run with Ollama?
36 of 63 have a tag in the Ollama library, shown in the table. The rest run through vLLM, SGLang, llama.cpp or Transformers using the commands their model pages list.
Do I need to accept a license before downloading?
For 6 models (Llama 4 Scout 17B-16E Instruct, Llama 4 Maverick 17B-128E Instruct, Cohere Command A (03-2025), Cohere Command R7B (12-2024), Cohere Aya Expanse 32B and AI21 Jamba Large 1.7) Hugging Face asks you to accept the license terms before the download starts. The others download directly.
Is this a benchmark or a ranking?
No. This page lists what exists, how big each download is and what the license allows. Tables sort by size or parameter count only; a bigger model is not automatically better for your task. For quality comparisons see https://www.developersdigest.tech/benchmarks.
What does GB mean on this page?
Decimal gigabytes: 1 GB is 10^9 bytes, the unit Hugging Face file sizes use. Operating systems that report binary units (GiB, 2^30 bytes) show the same file about 7% smaller. 85% headroom is applied on top of that.
Newsletter
New open weights, every week
One email per week. New model releases, open-weight drops and what they mean for what you build.
Keep exploring
Best local AI models
Our curated picks by hardware tier, with benchmarks and the reasoning.
Model API pricing and specs
First-party prices per 1M tokens when you would rather call an API.
AI benchmark leaderboards
Sourced scores, for when you need to compare quality.
Setup guides
Point your coding tools at the model you picked.