// category · 20 tools · 16 open source
Open Weight LLMs
Models whose weights you can download, inspect, fine tune and run yourself. Licences vary from fully permissive (Apache 2.0, MIT) to community licences with usage conditions. Always check before shipping.
What counts as an open weight model
An open weight model is one whose trained parameters you can download and run on your own hardware, usually from Hugging Face. That is not the same as open source: many open weight releases keep their training data and code private, and many come with licenses that restrict commercial use. Our open weight vs open source explainer walks through the difference and the main license types.
Licenses vary more than you might expect
Always read the model card before you build on a model. A few examples checked on Hugging Face on October 21, 2026:
- IBM Granite 4.0 H Small is released under Apache 2.0, a permissive license that allows commercial use.
- OpenAI's Whisper large v3 turbo speech model is MIT licensed.
- TII's Falcon H1 models use the separate Falcon LLM license rather than a standard open source license.
- Black Forest Labs' FLUX image models are mixed: FLUX.1 schnell and FLUX.2 klein 4B are Apache 2.0, while FLUX.1 dev, FLUX.2 dev and FLUX 3 Image Klein carry non commercial or community terms. Our FLUX open weights guide lists each one.
How to choose and run one
Start with memory. The weights, the context you plan to use and the quantization level decide whether a model fits on your GPU or Mac; our VRAM guide does the math, and GGUF quantization explained covers the file formats you will see. Then pick a runner. Ollama, LM Studio and llama.cpp suit one person on a laptop, while vLLM and SGLang suit a server with many users; see vLLM vs llama.cpp vs SGLang and Ollama vs llama.cpp vs LM Studio.
For a shortlist of models that run well at home, read open source LLM models to run locally. We also publish step by step guides for individual releases, such as Gemma 4, gpt-oss and Qwen 3.8. Each tool page below links to the official site and, where there is one, the GitHub repository, so you can check the current license and versions yourself.
Mistral AImistral.aiParis based lab publishing both open weight and commercial models, from small edge models to mixture of experts flagships.Open Weight ModelsFreemium★ 10.8k
FLUXbfl.aiBlack Forest Labs image models. Open weights include FLUX 3 Image Klein, FLUX.2 dev and klein, and FLUX.1 schnell and dev, under Apache 2.0, FLUX Kommunity and non commercial licenses.Open Weight ModelsFreemium★ 26k
Bonsai (Prism ML)prismml.comPrism ML's family of 1 bit and ternary open weight models, including Bonsai 2 27B, a Qwen3.8-27B derivative that fits in about 6 to 7 GB and runs on Prism's llama.cpp fork.Open Weight ModelsOpen source★ 3.5k
LFM2 / LFM2.5 (Liquid AI)liquid.aiLiquid AI's LFM2 and LFM2.5 open weight edge models, including LFM2.5-24B-A2B (24B MoE, ~2.45B active, 131K context) with official GGUF, MLX, vLLM and SGLang builds.Open Weight ModelsSee siteHN 19
Whispergithub.comOpenAI's open speech recognition model trained on a large multilingual dataset, widely used for transcription and translation.Open Weight ModelsOpen source★ 111.4kDeepSeek-V3github.comLarge mixture of experts model from DeepSeek whose technical report detailed efficient training tricks such as multihead latent attention.Open Weight ModelsOpen source★ 104.6kDeepSeek-R1github.comOpen reasoning model trained with large scale reinforcement learning, released with distilled smaller versions and an MIT licence.Open Weight ModelsOpen source★ 92k
Qwenqwenlm.github.ioAlibaba's Qwen model series, spanning dense and MoE language models plus vision, audio and coding variants, mostly under permissive licences.Open Weight ModelsOpen source★ 27.7k
Gemmadeepmind.googleGoogle DeepMind's lightweight open models built from Gemini research, sized to run on laptops, phones and single GPUs.Open Weight ModelsOpen source★ 5.8kgpt-ossgithub.comOpenAI's open weight reasoning models released under Apache 2.0, designed for local and on premise deployment.Open Weight ModelsOpen source★ 20.5k
MOSS-Image (OpenMOSS)openmoss.github.ioOpenMOSS's Apache 2.0 open weight text to image model: a 5B diffusion transformer with a Qwen3.5-9B text encoder, strong at long text and JSON layout control.Open Weight ModelsOpen source★ 2.1k
Llamallama.comMeta's family of open weight language models, released with weights for research and commercial use under the Llama community licence.Open Weight ModelsOpen source★ 7.7k
OLMoallenai.orgThe Allen Institute for AI's fully open language models, released with training data, code, checkpoints and logs.Open Weight ModelsOpen source★ 6.7kSmolLMgithub.comHugging Face's family of compact language models, published with their training recipes and datasets.Open Weight ModelsOpen source★ 3.9kPythiagithub.comEleutherAI suite of models trained on identical data in a fixed order, with many checkpoints for studying how training shapes behaviour.Open Weight ModelsOpen source★ 3k
IBM Graniteibm.comIBM's Apache licensed open models aimed at enterprise tasks such as RAG, code and tool calling.Open Weight ModelsOpen source★ 147
DeepSeek-V4.1-Flashhuggingface.coDeepSeek's MIT licensed open weight mixture of experts model: 552B backbone plus 196.6B of Engram lookup tables, image input and a 1M token context.Open Weight ModelsOpen source
Falconfalconllm.tii.aeModel family from Abu Dhabi's Technology Innovation Institute, including transformer and state space variants.Open Weight ModelsOpen source
Muse Spark Open (Meta)research.meta.aiMeta's open weight multimodal MoE model (816B total, ~42B active, 1M context) for coding and agents, with BF16 and NVFP4 weights plus community GGUF and MLX builds.Open Weight ModelsSee site
Recursive Flintrecursive.comRecursive's Apache 2.0 open weight coding model: a mixture of experts with 39.8B total and about 2.5B active parameters, a 262K context and tool calls.Open Weight ModelsOpen source// guides
Guides to open weight LLMs
Open Source LLM Models to Run Locally
Open source LLM models to run locally: match Qwen3, Gemma 4, gpt-oss and DeepSeek R1 distill tags to your VRAM, with Ollama setup and licence caveats.
Open Weight vs Open Source AI: The Difference
Open weight vs open source AI explained: what the OSI definition requires, which popular models are which, and a licence table checked against model cards.
SGLang vs vLLM on DGX Spark: Which to Use?
SGLang vs vLLM on DGX Spark compared on setup, memory, NVFP4 models, speed and clustering, so you can pick the right server for your GB10 desktop.
How Much VRAM Do I Need for an LLM?
How much VRAM do you need for an LLM? A simple formula for weights plus KV cache, worked examples from real model configs, and what fits in 8, 16, 24 or 32 GB.
How to Run DeepSeek Locally with Ollama
Run DeepSeek Locally with Ollama: pick R1 distill tags by download size, set context length, handle thinking output, and know when 671B V3 is server only.
How to Use Local LLM in VS Code
Wire a local LLM into VS Code with Ollama: official Chat extension, Continue.dev for tab complete, model picks by VRAM, and privacy tips for offline coding.
How to Quantize LLM Models
Practical how to quantize LLM models: llama.cpp convert then Q4_K_M, Hugging Face bitsandbytes 4 bit loads, Ollama GGUF import, and when to use each path.
KV Cache Quantization: Fit Longer Context in VRAM
KV cache quantization explained: how q8_0, q4_0 and FP8 caches cut memory, the settings for llama.cpp, Ollama, LM Studio, vLLM and MLX, and when quality drops.
How to Run DeepSeek V4.1 Flash Locally
Can you run DeepSeek V4.1 Flash locally? Real RAM, VRAM and disk needs by GGUF, Strata on one 12 GB GPU, the llama.cpp PR, vLLM, Ollama and DGX Spark.
SmolLM vs Llama: Which 3B Model Should You Run?
SmolLM vs Llama compared on benchmarks, licence, context, languages, memory and local setup, so you can choose between SmolLM3 and Llama 3.2 for your project.
SmolLM vs Gemma: Which Small Model to Use?
SmolLM vs Gemma compared on size, licence, context, languages, vision and audio, local setup and fine tuning, so you can pick the right small model in 2026.
Best LLM Pretraining Dataset in 2026: Top Picks
The best LLM pretraining dataset depends on your goal. FineWeb-Edu for small models, DCLM-Baseline and Nemotron-CC for scale, Common Pile for licensed text.
LLM Distillation vs Quantization: Key Differences
LLM distillation vs quantization: distillation trains a smaller student model, quantization stores the same weights in fewer bits. When to use each, or both.
Best Local LLM for Image Generation in 2026
The best local LLM for image generation is really an open image model. Our picks by GPU and license: FLUX 3 Klein, Z-Image Turbo, Qwen-Image 2.1 and more.
How to Run Bonsai 2 27B Locally (Ternary GGUF)
Run Prism ML's ternary Bonsai 2 27B locally: why stock llama.cpp, Ollama and LM Studio can't load it, PQ2_0 vs PTQ1_0, memory needs, setup steps and settings.
SGLang vs vLLM vs llama.cpp: Which to Use?
SGLang vs vLLM vs llama.cpp compared: what SGLang is, how RadixAttention works, Docker setup, independent benchmarks, hardware and when to pick each.
Ollama vs llama.cpp vs LM Studio: 2026 Guide
Ollama vs llama.cpp vs LM Studio compared for 2026: what each one is, ports, licences, Mac and GPU support, the context length trap, and which to pick.
Train a Small LLM From Scratch (nanochat, Megatron-LM)
Train a small LLM from scratch with nanochat, and learn when NVIDIA Megatron-LM fits: model sizes, budgets, datasets, compute math and multi GPU parallelism.
Speculative Decoding vs MTP: Faster LLM Output
Speculative decoding vs MTP explained: how draft models, EAGLE-3, MTP heads and DFlash speed up LLM output, and how to enable each in llama.cpp, vLLM or SGLang.
MLX vs GGUF on Mac: Which Should You Use?
MLX vs GGUF on a Mac: what each format is, how Ollama and LM Studio use them, speed tradeoffs, memory limits, fine tuning with mlx-lm, and which to download.
How to Run ComfyUI on Mac (Apple Silicon)
How to run ComfyUI on Mac: install Comfy Desktop or set it up by hand with PyTorch and Metal, where model files go, what fits in memory and how to fix errors.
Recursive Flint: Run the 40B Coding Model Locally
Recursive Flint is an Apache 2.0 coding model with 40B total and 2.5B active parameters. How to run Flint locally with Ollama, llama.cpp, rec or Unsloth.
MOSS-Image: Run It Locally in ComfyUI
MOSS-Image is an Apache 2.0 open weight 5B text to image model built for text and layouts. How to run MOSS-Image in ComfyUI, file sizes and best settings.
FLUX Open Weights: Every Model and Its License
FLUX open weights explained: every Black Forest Labs model you can download, its license, file size and VRAM, plus how to run FLUX 3 Klein in ComfyUI.