Open-source LLMs explained: what they mean for your brand's visibility

Open-source LLMs power third-party apps

Alibaba's Qwen and Meta's Llama sit under the largest number of third-party apps. DeepSeek handles the most routed traffic. A Hugging Face report on open models found that developers had built 151,448 models derived from Qwen. That's 2.6 times the number built from Meta models. Google's models had 82,506 derivatives. A derivative is a copy of a base model, fine-tuned for one job, and it ships as a support bot or an internal assistant.

On OpenRouter, a service that sends app requests to different models, DeepSeek served 14.37 trillion tokens between November 2024 and November 2025. Qwen and Llama served less traffic, at 5.59 trillion and 3.96 trillion respectively. Llama also runs Meta AI inside Facebook and Instagram, as well as in WhatsApp, according to TechCrunch. Keep in mind that a chatbot on someone's website almost never tells you which base model it runs on.

Open weight vs open source

Open weight means you get the trained model file. Open source, by the formal definition, means you also get what you'd need to rebuild it. The Open Source Initiative released its Open Source AI Definition 1.0 at All Things Open 2024. It requires the complete training code plus enough detail about the training data to rebuild the system.

Ayah Bdeir, who leads AI strategy at Mozilla, said the definition asks for enough data information that a "skilled person can recreate a substantially equivalent system". Few big releases clear that bar. Meta's Llama 4 Community License says that any company with more than 700 million monthly active users needs Meta's permission. That clause rules Llama out of the formal definition.

So check the license before you trust the label. Apache 2.0 and MIT licenses let you use a model for almost anything. Custom "community" licenses add conditions.

Leading open-source model families

Qwen leads the open field in 2026 by downloads and derivatives. Hugging Face counted about 2.045 billion Qwen downloads on its Hub in 2026. Google's models had about 418 million downloads and Meta's had 227 million. Four families cover most of what deployers actually run.

Qwen from Alibaba

Alibaba has released more than 460 Qwen models in sizes from phone-scale to data-center scale. Those small sizes are a big reason developers fine-tune Qwen for narrow tasks. Because Qwen dominates the derivative count, it is the leading base-model choice for bots on retailers' sites.

Llama from Meta

Llama started the wave of open-weight models in 2023. Mark Zuckerberg, Meta's CEO, said the family hit 1 billion downloads in March 2025. It's still the open model that enterprises use most, according to Menlo Ventures. But Meta hasn't shipped a major release since Llama 4 in April 2025, and that's slowed its momentum.

DeepSeek

DeepSeek is a Chinese lab known for strong reasoning at a low price. Developers favor access through cheap hosted APIs over running it themselves. That pattern explains why it leads on routed tokens while it trails Qwen on downloads.

gpt-oss from OpenAI

OpenAI released gpt-oss-120b and gpt-oss-20b on August 5, 2025, under the Apache 2.0 license. These were its first open-weight models since GPT-2. OpenAI's move toward open models is partial because its flagship GPT models stay closed.

How closed models differ

A closed model keeps its weights on the vendor's servers, so you can only use it through the vendor's app or API. ChatGPT and Claude work this way, as does Gemini. The vendor controls updates and pricing, and it decides whether web search is switched on. With an open model, whoever runs it makes those calls.

The better choice between an open model and ChatGPT depends on your business needs. Open models win on cost per token and on control. Closed models still lead at the hardest reasoning tasks, and you don't have to run them yourself. One respondent quoted in Menlo Ventures' mid-2025 update said its team started with Llama and DeepSeek but "they couldn't keep up with the performance of closed-source over time." If you process large volumes of sensitive data, an open model makes sense. If you need top quality with no infrastructure, a closed API is the better fit.

Why companies self-host models

Companies self-host so their data never leaves their own servers, and because running a model yourself gets cheaper per token at high volume. A McKinsey survey run with the Mozilla Foundation found 63% of organizations were using an open source AI model, while the figure for the technology industry was 72%. Most of them run these models alongside proprietary tools.

The hardware bar has dropped. gpt-oss-120b runs on a single 80GB GPU, and its smaller sibling fits in 16GB. Costs still depend on traffic, though, and on whether you have engineers to keep the system running.

A self-hosted assistant inside a bank or a patient bot at a hospital runs with no internet access at all. Everything it says about your company comes from training.

Training cutoffs freeze brand knowledge

Without web search, an open model knows your brand only as the public web described it before its training cutoff. Meta's model card for Llama 3.1 lists a December 2023 cutoff, and the model came out in July 2024. Llama 4 Maverick moved the cutoff to August 2024.

The weights never update. A company that put Llama 3.1 into its support tool in 2024 still runs that same file today. If you launched a product or rebranded after December 2023, that bot doesn't know. It is also unaware of price changes after that date. It also describes your old positioning with full confidence.

What the model does remember depends on how often and how consistently your brand appeared online before the cutoff. A brand with a handful of scattered mentions gets a vague description or a wrong one.

Retrieval depends on the deployer

An open model doesn't search the web on its own. The company that deploys it decides whether to connect a search tool or a document index, a setup called retrieval-augmented generation (RAG). Llama 3.1's prompt format includes slots for tools such as Brave Search, but the developer has to switch them on.

A shopping assistant pulls live search results, while a software vendor's help bot indexes only its own documentation and never sees your site. The base model can be identical in both cases, yet the answers about you differ.

When retrieval is on, your current pages count. The bot has to reach and read pages whose facts are stated clearly enough to quote.

What it means for brands

Your brand's visibility in open models has two layers, and you influence each one differently. The first is training memory. It comes from the public web before a cutoff, so it changes only when deployers move to newer models. The second is retrieval. It comes from pages the deployer's search tool can reach today.

Keep facts about your company consistent across your own site and the third-party sources that future training sets will scrape. Pages written in plain, quotable statements also hold up better when a retrieval system pulls a fragment out of context.

A thousand apps built on Qwen can give a thousand slightly different answers about you.

Auditing brand presence

A brand audit for open models compares what the models say about you with what's true, both with web search and without it. Closed assistants get audited the same way, but open models add the frozen-memory layer. Document your findings on:

• What base models such as Llama, Qwen and DeepSeek say about your brand with web search switched off
• Whether that training-memory answer reflects your current products and prices
• Whether AI crawlers and search tools can reach and read your key pages
• How consistently third-party sites describe your company's core facts
• How the closed assistants (ChatGPT, Gemini, Claude) answer the same prompts

The gap between the no-search answer and the with-search answer tells you which layer to fix. If the no-search answer is wrong, the problem sits in your older web footprint. If the with-search answer is also wrong, start with your pages.

When this matters for you

This matters most if your brand changed after 2023 or if people in your market use assistants built by third parties. Start by testing what base open models say about you with search turned off, because that answer sits unchanged inside apps for years. A wrong answer there won't fix itself. Your current pages and consistent third-party mentions are the only route into the next training run of any open source LLM. Snoika tracks how brands appear in Google and AI assistants and audits websites. It also builds researched content with human review. To see where your brand stands, check your AI visibility free.