Open Source LLM Releases Peaked in 2024: The Data and What It Means

Open source LLM releases peaked at 60% of all model launches in 2024 and have fallen every year since, to 38.3% in 2026. And the geography behind that number is stark: 85.9% of releases from Chinese labs ship open or partial weights, against 34.0% from everyone else.

"We will just use open models" has become the standard fallback in a lot of technology strategies. It answers vendor lock-in, price rises, data residency, and the risk that your provider deprecates the model you built on.

It is a reasonable plan. It also has a trend running against it and a geography attached to it, and almost nobody presenting it has checked either.

The share peaked two years ago

I classified all 261 significant model releases between 2018 and 2026 by weights licence – open, partial or proprietary. As a share of each year's total releases:

Two things are true here, and they point in different directions.

More open models are being released than ever. The total release count grew sharply, so the absolute number is up.

A smaller proportion of the field is open. The frontier is drifting proprietary.

If you care about how many open models exist, the picture is good. If you care about whether the best available model will be open in three years, the trend is against you. That is why this argument never resolves: both sides are quoting a real number.

The geography will surprise your board

Split the same data by where the lab is and the picture changes character entirely.

Chinese labs accounted for 27.2% of 2026 releases so far, and at 85.9% open they are carrying the open-weight ecosystem almost single-handedly.

So "we will just use open models" is not a neutral technical choice any more. It is, increasingly, a decision to run on weights from a particular set of labs. Whether that is acceptable depends on your sector, your customers, your regulator and your data. It can be a perfectly good answer.

What is not defensible is making that decision without noticing you have made it. I have sat in strategy sessions where the open-model fallback was presented as risk mitigation, and nobody in the room had checked whose models the fallback actually meant.

Three things that change how you should weigh it

The cost case has largely evaporated at the low end. The cheapest capable model fell from $0.60 to $0.05 per million output tokens between 2023 and 2026 – twelve-fold. If inference cost was your reason for going open, recompute it. Frontier pricing went the other way, up 25%, so the cost case survives at the top of the market.

The control case got stronger. Of 261 releases, 75.5% are already superseded or retired, and the median age of a model you can buy today is 60 days. The oldest survivor on the current shelf is an open-weight release at 520 days, and it survives precisely because open weights cannot be withdrawn.

That is the strongest argument for open weights in the whole dataset, and it has nothing to do with cost or capability. It is about permanence. A proprietary model can be deprecated on the vendor's timetable. An open one cannot.

Capital is paying for today's prices, not revenue. Disclosed funding across the private labs runs to roughly $356bn, against a funding-to-booked-revenue ratio of about 12.4x. Only 7 of 15 labs have an audited revenue figure at all.

If that capital gets more expensive, the first casualty is the free tier and the second is the $0.05 floor. Weights you already hold are immune to both. The first revenue line that does not depend on subscriptions has only just arrived: ChatGPT advertising reached a $1bn annualised run rate inside 200 days.

The counter-argument, stated properly

There is a real case that this trend reverses, and it deserves better than a dismissal.

Regulation is the strongest version of it. You cannot audit weights you cannot see, and policy is moving toward auditability. The EU Artificial Intelligence Act is the first comprehensive AI law of its kind and it treats free and open-source models differently from closed ones in several places. If that hardens into procurement rules, open share could climb again on policy alone, whatever the labs would prefer.

There is also a capability argument. The gap between the best open model and the best proprietary one has narrowed at several points in this dataset, and each time it narrows the case for paying frontier prices weakens for a large band of workloads. Annual denominators here are in the dozens, so one strong open release can move a year's figure noticeably.

What would change my mind, specifically: a full year in which open or partial share comes in above the 2025 figure of 47.6%. That has not happened since 2024.

Choosing an open source LLM: four questions in order

The open source LLM decision is usually argued on ideology and should be argued on four practical questions.

1. What happens if this model is withdrawn? If the answer is "we rebuild", proprietary is fine with a swap clause in the contract. If the answer is "we cannot", weights you hold are worth paying for. This question should dominate and it usually gets asked last.

2. Whose weights, specifically? Given 85.9% against 34.0%, "open" narrows your realistic options far more than the word suggests. Write the actual candidate list down before deciding, and take it to whoever owns compliance.

3. Is the cost case still real? At a $0.05 floor, for many workloads it is not. Work out your actual inference spend at current prices before assuming self-hosting saves money. You may be buying operational burden to avoid a bill that already fell twelve-fold.

4. What is your exposure if capital tightens? If your economics depend on today's frontier pricing or a free tier, held weights are insurance. If they do not, they are overhead.

One more factor now sits alongside these. Seven models on the current shelf carry a category naming cyber capability, and seven were never generally released at all. The first model graded critical on cyber capability by its own maker arrived this year. Restriction is a new axis in this market, and it applies differently to weights you hold than to an API you call.

Questions people ask about open source LLMs

These are the questions searched most often alongside open source LLM, answered from the release data above.

What is an open source LLM? A large language model whose weights are published, so you can download, run and modify it yourself rather than only calling it through someone's API. In this dataset the licence field distinguishes open, partial and proprietary, because many releases sit in between.

What are open weight AI models? The same thing, described more accurately. "Open weights" means the trained parameters are available. It does not always mean an open source licence in the software sense, and it rarely means the training data is published.

What is the best open source LLM? There is no durable answer at a release every 2.98 days. The useful question is which open models meet your requirements on licence, language, context and hardware, then test two of them on your own task and keep the test.

Are open source LLMs cheaper? Less often than people assume. The cheapest hosted model now costs $0.05 per million output tokens, twelve times less than in 2023, so self-hosting frequently costs more once you count engineering time.

Is open source AI declining? Not in absolute terms – more open models ship each year. As a share of all releases it has fallen from 60.0% in 2024 to 38.3% in 2026, and that share is heavily concentrated in Chinese labs.

---

Weights-licence figures: a workbook of 261 model releases across 15 labs, 11 June 2018 to 3 September 2026, compiled from vendor documentation, launch posts, API release notes and Wikipedia, captured 7 September 2026. Every figure was computed once and independently re-derived through a second code path. Full method in the report, How AI Changed the Way People Search.