Open source LLMs 2026 — the best freely downloadable AI models

Open source LLMs 2026 — the best freely downloadable AI models

Most people know ChatGPT, Claude and Gemini. Alongside them sits a different kind of AI: models anyone can download and run on their own, with no subscription. They are called open source models, and over the past year they have become surprisingly good. Here is the current ranking, why almost all of them come from China, and what that means for you.

What is an open source LLM?

A normal AI service works like a subscription. You send your question to OpenAI or Google, they run it on their computers and send the answer back. You pay per use, and your text passes through their servers.

An open source LLM — a language model with open source weights — is instead a file you get to download. It can run on a powerful computer at your office, or on a server you rent in Europe. Do that, and your information never leaves the building.

Three things make this interesting for a smaller company:

• No subscription. You pay for the computing power, not per question. If you run the same task thousands of times a month, the difference is large.
• Your data stays with you. Useful if you handle personal data, contracts or customer records.
• Nobody can pull the rug out. A model you have already downloaded cannot suddenly get more expensive, change or be discontinued. It is the same model in three years.

The catch is that someone has to set it up and keep it running. That is why most companies still start with a normal subscription — and only look this way once volume or privacy requirements justify it.

One thing that is often misunderstood: open source here usually means the model file itself is free to use. That is not always the same as the training data being public, or the licence allowing absolutely everything. Read the licence before you build something commercial on a model.

The best open source models right now

Independent site Artificial Analysis puts every model through the same tasks and gives it a combined score for how smart it is. Here is the ranking among freely downloadable models as of September 2026. You do not need to read every line — the pattern is the point.

• GLM-5.3 (Z.ai, China) — score 45, around 2 dollars per completed task. Best of all open models.
• Kimi K3 (Moonshot AI, China) — score 44, around 2 dollars.
• GLM-5.3-Flash (Z.ai, China) — score 42, around 25 cents.
• Qwen3.8 (Alibaba, China) — score 40, around 2.15 dollars.
• DeepSeek V4.1 Flash (DeepSeek, China) — score 40, around 27 cents.
• DeepSeek V4 Pro (DeepSeek, China) — score 36, around 67 cents.
• Nemotron 3 Ultra (NVIDIA, USA) — score 23. The best the West has, and far behind.

See the full list and updated figures at Artificial Analysis

Two things stand out. The first is the price gap: GLM-5.3-Flash scores almost as high as its big sibling but costs an eighth as much. Picking the most expensive model is rarely the right answer. The second is that the entire top of the list comes from one country.

The seven names you will run into

The ranking above measures one thing: how smart the model is. There are fewer families behind those rows than there are rows, and these are the ones you will see in tools, cloud services and articles.

• DeepSeek — the Chinese company that broke through in 2025 by matching the American models at a fraction of the cost. It comes in two tracks: a general model and one built to reason its way to an answer step by step.
• Qwen (Alibaba) — the broadest of them all. Handles over 200 languages, takes both text and images, and comes in everything from small versions that run on a laptop to giants. Free licence on most versions.
• GLM (Z.ai) — the model currently topping the list. Built for longer assignments where the AI has to hold many steps together, such as programming.
• Kimi (Moonshot AI) — specialised in working with images and text at the same time, and in splitting a large assignment into smaller pieces that run in parallel.
• MiniMax — aimed at everyday office work and programming, and at solving it in as few steps as possible. Cheap to run.
• Llama (Meta) — the best known Western open model, and the one with the most ready-made guides and tooling around it. On performance it has fallen behind.
• Gemma (Google) — Google's little sister to Gemini. The point is not top results but that it is small enough to run on your own hardware, even on a phone.

If you have to pick one today with the least possible friction: Qwen has the broadest tool support, GLM the best results, and Gemma the lowest hardware requirements.

Almost the entire top is Chinese — and many models are distilled

Six of the seven best open models come from Chinese companies. The best Western open model, NVIDIA's Nemotron, scores 23 against GLM-5.3's 45. Meta's Llama and Google's Gemma sit further down than that.

Part of the explanation is strategic. The American frontier models are kept closed because they are the product — OpenAI and Anthropic make money on subscriptions. Chinese labs instead use open releases to win users and developers quickly.

But a large part of the explanation is technical, and it is called distillation.

What does distillation mean?

Distillation is a technique where a small model is apprenticed to a large one. The large model answers hundreds of thousands of questions, and the small one is trained to imitate those answers as closely as possible. Think of an apprentice following a craftsman for a year: they do not learn everything the master knows, but enough to do almost the same job — much faster and much cheaper.

The result is a model a fraction of the size but nearly as capable. That is why a Chinese model can cost a tenth of an American one and still keep up on most everyday tasks. It also explains why several of them exist in "Flash" or "Mini" versions: same knowledge, smaller package.

Distilling your own large model into a smaller one is entirely legitimate, and every major AI company does it. What is contested is when someone is suspected of having distilled a competitor's model by feeding it questions. OpenAI has accused DeepSeek of exactly that, but no evidence has been presented publicly.

Three things to consider with Chinese AI models

They are good and they are cheap. But there are things to be aware of before you let them into your business.

• They are censored. Ask a Chinese model about Tiananmen Square, the status of Taiwan or criticism of the government in Beijing, and you get evasive answers or none at all. For invoice handling and product copy that does not matter. For market intelligence, journalism or research it does. Test a couple of sensitive questions yourself before you decide.
• The app and the downloaded model are not the same thing. If you use DeepSeek's own app or cloud service, your text is sent to servers in China. If you instead download the model and run it yourself, or with a European provider, that does not apply. That difference is decisive for GDPR, and it is often missed in the discussion.
• The values come along. What the model was trained on, and how it was fine-tuned, reflects where it was built. It rarely shows in everyday tasks, but keep it in mind when the answer touches politics, history or social issues.

Worth saying plainly: the American models also have limits and values built in. The difference is that the censorship in the Chinese ones is state-motivated and considerably more predictable — it hits the same topics every time.

What does it cost to run a model yourself?

This is where most people are surprised, in both directions. Downloading the model costs nothing. The computing power does.

A small model like Gemma or a smaller Qwen runs on a modern office computer or a Mac with enough memory — free, entirely local, no internet needed. That goes further than many expect for summaries, templates and simpler questions.

The large frontier models need server hardware. In that case you rent capacity instead, preferably with a European provider, and pay by the hour or by usage. It is not always cheaper than a subscription — but you know where the data sits, and the price does not change because someone else decided it should.

What we would do

• Writing text, summarising meetings, asking questions? Stay in ChatGPT or Claude. Simpler, and the quality is worth the money.
• Got a task that runs thousands of times a month? Do the maths on an open model. That is where the calculation flips.
• Data that cannot leave the building? A model you run yourself is often the only route, and a European provider the easiest compromise.
• Start small. Take a task you already do, run it both ways for a month and compare results and cost. It is the only way to know whether it pays off for you.

If you want to read on, we have covered what DeepSeek is and the basics of what an LLM is. And if you want help deciding which model fits you — and building what will run it — we do that as an AI consultant.