From $1.50 usage-based🇫🇷
Multimodal model from Paris-based Mistral AI for agents and coding, via API and with open weights.
Best open-weight and local LLMs for business (October 2026)
This category covers language and embedding models whose weights are published, so companies can download them and run them on their own servers, in a private cloud or on laptops.
In this guide:24 toolsBefore you buyHow to choose open-weight and local LLMs for business (October 2026)Questions
24tools in this category
5countries with local strength
1can host data in the EU
How results are ranked: How we review
Results
Usage-based · Free trial🇨🇦
Mixture-of-experts model from Toronto-based Cohere with reasoning, image input and 48 languages, also with open weights.
From $20 per month · Free plan🇺🇸
Open reranker models from Mixedbread under Apache 2.0 for more than 100 languages and code, also available through the vendor's search API.
Free🇺🇸
IBM's open enterprise models in 3B, 8B and 30B sizes, under Apache 2.0 and with ISO 42001 certification.
Usage-based · Free trial🇩🇪
Reranker models from Berlin-based Jina AI, part of Elastic since 2025: multilingual, with long context and a multimodal variant.
Free🇨🇳
Open multilingual reranker model from the Beijing Academy of Artificial Intelligence under Apache 2.0, light enough for your own servers.
Free🇨🇳
Open embedding model from the Beijing Academy of AI for more than 100 languages, with dense, sparse and multi-vector retrieval in one model.
Free🇨🇳
DeepSeek's open multimodal mixture-of-experts model with one million tokens of context and a much smaller KV cache.
Free🇫🇷
Mistral AI's open coding model for software engineering agents, running with 24 billion parameters on an RTX 4090 or a 32 GB Mac.
Free🇺🇸
Hugging Face pretraining dataset of 1.3 trillion tokens of educational web pages, selected from Common Crawl by an AI quality classifier.
Free🇺🇸
Google DeepMind's open multimodal model family in five sizes, from E2B for mobile devices to 31B, with thinking modes.
Free🇺🇸
OpenAI's open-weight reasoning models in two sizes, 120b for one 80 GB GPU and 20b for 16 GB of memory, under Apache 2.0.
Free🇩🇪
Multilingual embedding models from Berlin-based Jina AI; the small version has 677 million parameters and 32,768 tokens of input.
Free🇫🇷
Mistral AI's open hybrid model that unifies instruct, reasoning and coding, with image input and a 256,000-token context.
Free🇨🇳
Alibaba's open coding model for coding agents with 80 billion parameters, only 3 billion of them active.
Free🇨🇳
Alibaba's current open-weight generation with image and video understanding, adjustable thinking and up to 1M tokens of context.
Free🇩🇪
Small open language model from the German OpenGPT-X project, trained for all 24 official EU languages.
Free🇺🇸
Open foundation model from Amazon for zero-shot time series forecasting, supporting multiple series and covariates.
Research paper🇺🇸
Meta FAIR technical report on DINOv3, a family of self-supervised vision models with high-quality dense features.
Research paper🇨🇳
Paper from Beijing (Tsinghua and ModelBest) on small language models with 1.2 and 2.4 billion parameters and scalable training strategies.
FreeAvailable internationally
Massive benchmark and toolbox for embedding and retrieval models, with more than 1,000 tasks and a public leaderboard on Hugging Face.
Free🇺🇸
Open vision-language-action model from NVIDIA for manipulation tasks on humanoid and other robots, with fine-tuning on your own data.
Free🇺🇸
Open foundation model from Google Research for time series forecasting with quantile output, also available in BigQuery.
Free🇺🇸
Open-source tool for faster fine-tuning of language models with less GPU memory, including LoRA, QLoRA and reinforcement learning.
How to choose open-weight and local LLMs for business (October 2026)
This category covers language and embedding models whose weights are published, so companies can download them and run them on their own servers, in a private cloud or on laptops. It is split into general purpose models, coding models for software engineering agents, embedding models that power semantic search and RAG, and small models that run on a single GPU or on end devices.
Companies choose open models when data must not leave their own infrastructure, when they want predictable costs at high volume or when they need to fine-tune a model on internal examples. Typical uses are internal chat assistants, document extraction, code assistants that never send source code outside and knowledge search where an embedding model turns documents into vectors.
Open does not always mean free for business use. Read the license: Apache 2.0 and MIT allow commercial use, licenses marked non-commercial require a separate agreement, and some vendor licenses such as the Qwen Community License restrict certain commercial offerings. Within one model family, sizes can carry different licenses. Plan the operation, since your team becomes responsible for GPU capacity, security updates, monitoring and model upgrades. Check the hardware a model needs, test quality in your languages and document which version you run, so you can reproduce results and react when the maintainer releases a new version.
Questions about open-weight and local LLMs for business (October 2026)
Are open-weight models free for commercial use?
Many are, for example under Apache 2.0 or MIT. Some are published under non-commercial licenses such as CC BY-NC 4.0, and some vendor licenses limit specific commercial uses, such as offering the model as a service to third parties. Always check the license of the exact version and size you deploy.
What hardware do I need to run a model locally?
It depends on size. Small models with a few billion parameters run on laptops or a single consumer GPU, while large mixture-of-experts models need one or more data-center GPUs. The model card usually states the requirements.
Does running an open model locally solve data protection questions?
It keeps prompts and documents inside your infrastructure, which removes the transfer to an external provider. You still need access controls, logging rules and a clear purpose for processing personal data.
Last updated: