Open source vs proprietary AI: what is the difference for you?

Proprietary AI can only be used through its maker’s app or API, whereas an open model can be downloaded and run on your own computer or server. For you, the difference comes down to four things: privacy, cost, performance and freedom of use, which we compare here using the site’s data as of October 22, 2026.

Open source, open weights, proprietary: three ideas to separate

An AI model is essentially a huge set of numbers: its parameters, also called weights, set during training. The whole debate is about what the maker agrees to publish.

• Proprietary (or closed): the weights stay on the maker’s servers. You use the model through an assistant such as ChatGPT or Claude, or through an API, a doorway that lets programs talk to the model. GPT-6, Claude and Gemini work this way.
• Open weights: the weights are published and downloadable, under a more or less permissive licence. This is an open weights model. Most DeepSeek, Qwen, Kimi and GLM models fall into this group.
• Open source in the strict sense: on top of the weights, the maker publishes the training code and enough information about the data to rebuild the model.

In practice, very few models tick every box. Makers and the press often use “open source” loosely. When you read “open source model”, it usually means “open weights model”. A few players go further: Ai2 publishes the data, code and intermediate training checkpoints of its OLMo models, and NVIDIA releases its Nemotron models with their datasets and recipes, under its own licence.

The licence: the detail that changes everything

An open model still comes with conditions. According to the site’s maker profiles:

Maker	Most common licence for open models	Commercial use
Mistral AI	Apache 2.0 (some under a modified MIT)	Yes
Alibaba (Qwen)	Apache 2.0 for most	Yes
DeepSeek	MIT	Yes
Zhipu AI (GLM)	MIT	Yes
Moonshot AI (Kimi)	Modified MIT	Yes, with added clauses
IBM (Granite)	Apache 2.0	Yes
Meta (Muse Spark Open)	Meta’s own licence	Check first
NVIDIA (Nemotron)	NVIDIA Open Model License	Yes

Apache 2.0 and MIT are short, very permissive licences. In-house licences may restrict certain uses or add obligations. The rule is simple: read the licence file of the exact model you download, because it can change from one version to the next.

Performance: the gap with closed models

Closed models are still ahead. In the weekly ranking, built from Epoch AI Benchmarking Hub tests as of October 22, 2026, the top ten spots all go to models from Anthropic, Google DeepMind and OpenAI. The first open weights model in the top 20 is Kimi K2.6, from Moonshot AI, in 19th place with an index of 84.4 out of 100.

The chart below compares the Epoch index of several recent open models with three leading closed models.

Capability index (out of 100)

• Claude Sonnet 5.595.9
• Gemini 4 Argon95.3
• GPT-6 Astra94.7
• DeepSeek V4.1 Flash 101790.6
• Kimi K2.684.4
• Kimi K380.7
• GLM-5.378.8
• Muse Spark Open75.2
• Qwen3.6 27B72.2

A benchmark measures specific tasks, such as science questions or code. It does not tell you whether a model suits your needs. To draft an email, summarise a report or sort documents, a lower-ranked model may do the job perfectly well. Our article on AI benchmarks and their pitfalls helps you read these numbers with a critical eye.

The numbers: price, context, openness

Here is a selection of open weights models recorded by models.dev on October 22, 2026, with one closed model for reference. Prices are for API access, in US dollars per million tokens, the small chunks of text that models process.

Model	Developer	Release	Context (tokens)	Input ($/M tokens)	Output ($/M tokens)	Open weights
DeepSeek V4.1 Flash 1017	DeepSeek	October 17, 2026	1.05 M	$0.15	$0.60	Yes
Kimi K2.6	Moonshot AI	April 20, 2026	262 k	$0.95	$4	Yes
GLM-5.3	Zhipu AI (Z.ai)	August 14, 2026	1 M	$1.2	$4.18	Yes
Muse Spark Open	Meta	October 15, 2026	1.05 M	$1.25	$4.25	Yes
Qwen3.6 27B	Alibaba Cloud (Qwen)	April 22, 2026	262 k	$0.33	$2.7	Yes
Gemma 4 31B IT	Google DeepMind	April 2, 2026	262 k	$0.14	$0.40	Yes
Mistral Small 4	Mistral AI	March 16, 2026	262 k	$0.16	$0.63	Yes
Nemotron 3.5 Super VL 120B A12B	NVIDIA	October 15, 2026	1.05 M	$0.15	$0.50	Yes
Claude Sonnet 5.5	Anthropic	September 28, 2026	1 M	$2	$10	No

Two lessons stand out. First, price: DeepSeek V4.1 Flash costs $0.15 for input and $0.60 for output per million tokens, against $2 and $10 for Claude Sonnet 5.5. Second, the context window, meaning how much text the model can read at once, often exceeds one million tokens for large open models, just as for closed ones. Smaller models such as Qwen3.6 27B have more modest windows, but they are also far easier to run yourself.

To dig deeper, our model comparison tool puts two models side by side, and the open source models page lists open models with their licence and size.

Privacy, control, lock-in: what really matters

The main argument for open models is not price but control.

• Your data stays with you if you run the model yourself: nothing is sent to the maker. Be careful: if you go through a third-party host, its terms apply.
• You can adapt the model through fine-tuning on your own examples, or shrink it with quantization so it fits on modest hardware.
• You are not locked in: a downloaded model does not vanish if the vendor changes its prices or terms, or retires a version.

Proprietary models, on the other hand, offer convenience: no server to manage, automatic updates, built-in features (web search, voice, agent mode) and, for now, the best raw performance. Running a large open model takes expensive hardware, skills and maintenance time. Our guide to running AI locally covers the hardware you need.

Openness also has a geopolitical side: many of today’s strongest open weights models come from China, while Mistral AI leads the European offer. Our article on European AI versus the US and China takes stock, and our overview of the companies that build AI profiles each player. The EU’s AI rules also reflect this distinction, with lighter duties for some models released under free licences, as explained in our EU AI Act explainer.

Check yourself A small firm wants to summarise client contracts without them leaving its network. Which option should it favour?

An open weights model installed on an internal server, under a licence that allows commercial use (Apache 2.0 or MIT, for example). The contracts are then never sent to an outside service. The firm will need to plan for hardware, security and human review of the summaries.

Try an open model without installing anything

You do not need a server to get a feel for open models. Several consumer assistants already offer them in their interface, according to the site’s data as of October 22, 2026:

• Duck.ai, DuckDuckGo’s chat, usable without an account, includes gpt-oss-120b and Gemma 4 31B among its models;
• Lumo, Proton’s assistant, relies on models such as Qwen 3.5 and GLM 5.3;
• Brave Leo, built into the Brave browser, gives access to Llama, Qwen, DeepSeek and Mistral models.

It is an easy way to compare their answers with those of your usual assistant, on your own questions, before considering an installation.

Verdict: which type of model, for whom?

Your profile	Suggested option	Why
Individual, everyday use	Proprietary assistant	Simple, free basic tier, built-in features
Tech-curious user	Small open model run locally	Learn, test offline, no subscription
Developer, high-volume project	Open model through an API	Price per token often much lower
Business with sensitive data	Open model hosted in-house	Data under control, no lock-in
Need for top performance	Leading proprietary model	Top of the ranking on October 22, 2026

There is no single “best” choice: many organisations combine both, a closed model for the hardest tasks and an open model for volume or sensitive data.