Open-source AI models explained: Llama, Gemma, DeepSeek and the rest

What open-weight AI models are, how the Llama, Gemma, DeepSeek, Qwen, Mistral and gpt-oss licences differ, and when a UK business should use one.

An open-weight AI model is one whose trained parameters, the “weights”, are published so anyone can download the model and run it on their own computers or cloud account. Llama from Meta, Gemma from Google, DeepSeek, Qwen from Alibaba, Mistral’s smaller models and OpenAI’s gpt-oss are the best-known families. Closed models, such as the ones behind ChatGPT, Claude and Gemini, can only be used through the company’s own apps and API.

“Open source” is often used loosely for all of these, but the licences differ a great deal. Some are as permissive as ordinary open-source software. Others come with conditions on naming, attribution, size of business or where you are based. For a UK business the licence, not the label, decides what you can do.

  • Open weights let you run a model on infrastructure you control, which can help with data protection, cost at scale and avoiding lock-in.
  • Gemma 4, Qwen3.5 and gpt-oss use the Apache 2.0 licence and DeepSeek V4 uses MIT, both permissive. Llama 4 uses Meta’s own community licence with extra conditions.
  • Among 11,895 developers in Stack Overflow’s 2026 survey, 18.9% used DeepSeek models, 10.1% Qwen and 9.8% Mistral, against 74.6% for Anthropic and 74.3% for OpenAI.
  • Running a model yourself means you also take on hosting, security, updates and quality checks. For most small firms, a hosted closed model is still simpler.
Developers using each provider's models, 2026
74.6%Anthropic
74.3%OpenAI
18.9%DeepSeek
10.1%Alibaba Qwen

Source: Stack Overflow Developer Survey 2026, 11,895 respondents

Open source, open weights and closed: the difference

With ordinary software, “open source” has a precise meaning: you get the source code and a licence that lets anyone use, study, change and share it. AI models don’t fit neatly. A model is mainly its weights, but to rebuild it from scratch you would also need the training data and code, which almost no company publishes in full.

The Open Source Initiative, which maintains the standard definition of open-source software, published its Open Source AI Definition 1.0 in October 2024. It expects the weights, the code and detailed information about the training data to be available under terms that allow free use and modification. Most well-known “open” models don’t meet that bar, which is why many people prefer the term open-weight.

In practice, three questions matter more than the label:

  • Can you download the weights? If not, it is a closed model, whatever the marketing says.
  • Can you use it commercially without asking permission? Permissive licences such as Apache 2.0 and MIT say yes. Custom licences may add conditions.
  • Are there conditions on what you build? Naming rules, attribution, acceptable-use policies or limits for very large companies.

How the main model licences compare

This comparison is based on each publisher’s licence and model pages, checked 22 November 2026, not on hands-on testing. Hahu has no affiliate or commercial relationship with any of these companies. Models and licences change often, so always read the licence attached to the exact model version you download.

Model family Publisher Licence Key conditions
Gemma 4 (April 2026, up to 31B parameters) Google Apache 2.0 Permissive. Google says these are the first Gemma models under an OSI-approved licence; earlier Gemma versions used Google’s own terms.
gpt-oss-120b and gpt-oss-20b OpenAI Apache 2.0 Permissive. Must be used with OpenAI’s “harmony” response format to work correctly.
Qwen3.5 Alibaba Apache 2.0 Permissive for the models we checked. Alibaba also sells closed Qwen models, so check each one.
DeepSeek V4 DeepSeek MIT Permissive. V4-Pro has 1.6 trillion parameters in total, with 49 billion active per token, and a one-million-token context.
Mistral Small 4 and Ministral 3 Mistral AI Apache 2.0 Permissive. Mistral Large 4 uses a separate Mistral Ecosystem Open Weight License, and Mistral’s “Premier” models are available only through its API.
Llama 4 (April 2025) Meta Llama 4 Community License Show “Built with Llama”; models built from it must have names starting with “Llama”; acceptable use policy applies; firms with over 700 million monthly users need Meta’s permission.

Two Llama details are worth knowing. Meta’s acceptable use policy withholds the licence rights for Llama 4’s multimodal models from individuals and companies based in the European Union. That restriction does not cover UK-based businesses, though it would matter if you have an EU parent company. And in April 2026 Meta released Muse Spark, its first model from Meta Superintelligence Labs, through the Meta AI app, meta.ai and a private API preview, with no downloadable weights. DeepLearning.AI’s newsletter The Batch described it as a pivot away from Meta’s open-weights Llama strategy. For now, Meta’s newest model is not open-weight.

How widely open models are used

Developers are the best-measured group. Stack Overflow’s 2026 Developer Survey asked which providers’ models people had used, and while the closed models from Anthropic, OpenAI and Google lead, open-weight families have a solid minority following.

Developers who used each provider's AI models, 2026

Respondents could pick several. Figures cover each provider's open and closed models together.

Source: Stack Overflow Developer Survey 2026, technology section, 11,895 respondents

View data
 Developers who used each provider's AI models, 2026
Anthropic74.6%
OpenAI74.3%
Google AI54.5%
Microsoft26.4%
DeepSeek18.9%
Alibaba Qwen10.1%
Mistral AI9.8%
xAI7.6%
Meta AI6.3%

Two caveats. The survey counts each provider once, so OpenAI’s figure includes both ChatGPT-style closed models and gpt-oss, and Google’s includes Gemini and Gemma. And developers are not typical business users. We know of no reliable UK figure for how many small businesses run open-weight models; government and ONS surveys ask about AI use in general, not which kind of model.

Meta’s 6.3% is striking for a company whose Llama models dominated open-weight AI a couple of years ago, and fits the shift in its strategy described above.

Why a business would choose an open-weight model

  • Control over data. Run the model on your own servers or a UK cloud region, and customer data never leaves infrastructure you control. That can simplify a data protection impact assessment, though it doesn’t remove the need for one. See our AI agents and UK GDPR checklist.
  • Cost at volume. With a closed API you pay per token, which we explain in what AI costs to run. With your own hosting you pay for computers whether they are busy or not. For steady, high-volume, fairly simple tasks such as classifying or extracting data, a smaller open model can work out cheaper. For light or unpredictable use, it usually doesn’t.
  • Customisation. You can fine-tune open weights on your own examples, which is useful for a fixed style or a narrow task.
  • No lock-in. The model you downloaded will still work the same way next year. Closed models get retired and replaced on the vendor’s timetable.
  • Offline and edge use. Small models run on laptops, phones and on-site hardware. OpenAI says gpt-oss-20b runs in 16GB of memory or less, and gpt-oss-120b fits on a single 80GB data-centre graphics card.

The downsides to weigh

  • You become the operator. Hosting, scaling, monitoring, security patches and upgrades are now your job or your developer’s. Our guide to what it costs to keep an app running applies here too.
  • Quality gap at the top. The largest open models are capable, but they need serious hardware. The small ones that run cheaply are noticeably weaker at complex reasoning than leading closed models.
  • Safety is on you. Hosted products include content filters and abuse monitoring. A downloaded model includes whatever the publisher built in, and fine-tuning can weaken it.
  • Licence obligations. Permissive licences still require you to keep copyright notices. Custom licences such as Llama’s need reading properly before you ship a product.
  • The hosted app is not the model. Using DeepSeek’s or another provider’s own app or API sends your data to that company, under its terms and jurisdiction. Running the same open weights on your own infrastructure does not. The ICO’s October 2026 update on its work with AI developers named DeepSeek among the companies it had engaged with.

There is no UK law specific to open-weight models; the usual rules on data protection, consumer law and intellectual property apply. In the EU, the AI Act gives providers of general-purpose models released under a free and open-source licence, with public weights, an exemption from some documentation duties under Article 53(2). They must still have a copyright policy and publish a summary of training content, and the exemption doesn’t apply to the most powerful models classed as having systemic risk. As a business using such a model, your duties depend on what you build with it, which we cover in UK AI regulation and the EU AI Act.

How to choose

Choosing between open and closed models

Step 1 of 6

Define the task

Write down the job, the volume per month and how good the answers must be.

Step 2 of 6

Start closed

Prototype with a hosted model to prove the task works before worrying about hosting.

Step 3 of 6

Check the data

Decide whether the data can go to a third-party API under a proper contract.

Step 4 of 6

Read the licence

Confirm the exact model version allows your commercial use, naming and attribution.

Step 5 of 6

Compare costs

Price API tokens against hosting at your real volume, including staff time.

Step 6 of 6

Test properly

Run both options on the same real examples and compare errors, not demos.

Our usual advice for a small firm: build the first version on a hosted model, design the code so the model can be swapped, and move to an open-weight model when volume, data rules or a customer contract give you a concrete reason. That keeps options open without running servers you don’t yet need. If the decision is part of a bigger build-or-buy question, see our guide to bespoke software or off-the-shelf. Open models also pair well with document search; our explainer on RAG and chatting with your documents shows how.

Common questions

Is open-source AI free?

The licence is usually free. Running the model is not: you pay for the computers, or for a hosting company that runs it for you, plus the time to set it up and maintain it.

Is an open model safer for customer data?

It can be, if you run it on infrastructure you control and secure it properly. Simply choosing an open model and using someone else’s hosted app gives you no data protection advantage.

Can we use Llama in a commercial product?

Generally yes, under the Llama 4 Community License, if you follow its conditions: show “Built with Llama”, start derived model names with “Llama”, follow the acceptable use policy and, in the unlikely case you exceed 700 million monthly users, get Meta’s permission.

Which open model is best?

There is no single answer, and rankings change every few months. Pick two or three candidates whose licence suits you and test them on your own task.

Where to start

If you are building a product and expect high volumes or strict data requirements, plan for model choice from day one. Our SaaS product development team designs systems that can run on hosted or open-weight models and switch between them as the economics change.

Sources

  • Stack Overflow Developer Survey 2026: Technology (October 2026, checked 22 November 2026)
  • Meta: Llama 4 Community License Agreement (effective 5 April 2025, checked 22 November 2026)
  • Google Open Source Blog: Gemma 4, expanding the Gemmaverse with Apache 2.0 (2 April 2026)
  • Meta: Llama 4 Acceptable Use Policy (checked 22 November 2026)
  • Hugging Face: deepseek-ai/DeepSeek-V4-Pro model card; Qwen/Qwen3.5-397B-A17B model card; openai/gpt-oss-120b model card (checked 22 November 2026)
  • Mistral AI documentation: Models overview (checked 22 November 2026)
  • Meta AI blog: Introducing Muse Spark, the first model from Meta Superintelligence Labs (8 April 2026)
  • DeepLearning.AI, The Batch: Life after Llama, with Muse Spark Meta pivots away from its open-weights Llama strategy (April 2026)
  • Open Source Initiative: The Open Source AI Definition 1.0 (October 2024)
  • EU AI Act, Article 53: Obligations for providers of general-purpose AI models
  • ICO: ICO secures changes from leading AI developers as scrutiny extends to AI agents (8 October 2026)

→Keep reading

Related guides.

All insights
Software reviews

Cursor vs GitHub Copilot vs Claude Code: which AI coding tool fits your team?

Cursor, GitHub Copilot and Claude Code compared on published prices, plans, usage limits and developer survey data, with a best fit verdict for UK… Read
AI explained

RAG explained: how “chat with your documents” AI really works

Retrieval-augmented generation lets AI answer from your own files. How RAG works, what it costs, why it still gets things wrong and when to… Read
SaaS & CRM

How to build a SaaS product: from idea to first paying customer

A practical route from SaaS idea to first paying customer: validation, scope, the parts every product needs, security basics, launch and what drives cost. Read

→Get in touch

Got a question? Ask us.

Tell us what you’re trying to do. Every project starts with a free audit: someone who does the work replies with a few ideas, an honest view and a written plan with prices.

What happens next

  1. 1
    You send the formIt takes about two minutes.
  2. 2
    We do our homeworkWe look at your website and your market.
  3. 3
    You get a short planWhat we’d do first, with prices. Yours to keep.
  • Free, with no obligation
  • We won’t add you to a mailing list

We only use your details to reply to you. See our privacy policy.