Build1 publisherNot yet confirmed elsewhere2 min readPublished
Percona CEO asks the industry to stop calling open-weight models open source
Percona CEO Peter Farkas says open-weight models lack open source's freedoms, even as they carried 56% of Vercel AI Gateway tokens in August. For teams choosing a model, the license and what was actually released now belong in the review.
The Engineer · Build desk
What happened
- On OpenRouter, open-weight models took 60% of US-originating token consumption in August, and Chinese-developed models made up the majority of it.
- Farkas named DeepSeek as an example of the confusion, since its models are widely called open source though it released weights, not everything needed to reproduce them.
- Xiaomi livestreamed nearly a week of MiMo-V2.6 reinforcement-learning training and released over 7,000 RL task environments with its training code and documentation.
- The Open Source Initiative published its first Open Source AI Definition in 2024, setting criteria around freedoms such as the freedom to use.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- decision I think model selection now needs a recorded line for what shipped besides the weights, because releases sold as open differ widely in what they expose.
- constraint A team holding weights alone can host the model but cannot reproduce it from scratch, so a full retrain still depends on the lab that trained it.
- precedent Farkas warns that a looser meaning accepted for AI would weaken the definition of open source for ordinary software as well.
An open-weight release is a file of numbers. The weights are the parameters produced in training, and they encode the patterns the model learned [4]. With them, a team can download the model and run it on its own infrastructure, even if the creator never released the ingredients or the process that produced it [5]. "You don't have the source code, you don't have the training data, you only have the output of these two," Farkas said. "So why would we call this open source in the first place?" [6]
James Landay, director of Stanford's Institute for Human-Centered AI, drew the same line in an HAI article in August [7]. "Open weights answer 'can I run this?'" Landay said. "Open source answers 'can I trust this, improve it, and build the next thing on top of it?'" [8] He argues that the major labs mostly answer the first question and fall well short of the second [9].
Farkas, who co-created FerretDB before running Percona, made his case at Open Source Summit Europe in Prague [1]. He still rates the weights highly. "Are open weights a bad thing? No, open weights are great," he said. "You can run your models in your own environment, you can experiment with them, and if you understand the risks, you can also use it in production. The problem is when open weights are positioned as, 'hey, this is as good as open source'." [10]
The usage figures deserve the same care as a benchmark table. Both gateway shares are above half [19]. They count tokens routed through Vercel's AI Gateway and US-originating consumption on OpenRouter [2][3]. They describe those gateways' customers [2][3]. For either share to transfer to a particular team, that team's traffic mix would have to resemble a model gateway's customer base.
The AI industry took up the language of open source while some of its best-known "open" models deliver only part of what the term promised [18]. For an operator, the gap shows up in the license. Farkas said companies are going to get away with calling something Apache 2.0 even when, for example, it can't be used in the European Union [15]. The reported token shares are split by open-weight status, not by license [2][3].
Xiaomi's MiMo-V2.6 release is the most careful work in the reporting. It gives outsiders a view of how the models were fine-tuned, and it shows how far apart "open" releases can sit in what they expose about how a model was built [12]. Whether it is enough to count as open source AI is a separate question, The New Stack wrote [12].
What to watch
- Whether the OSI revises its Open Source AI Definition and says which high-volume models meet it.
- Whether later Vercel or OpenRouter token reports split open-weight traffic by license.
- Whether other labs follow Xiaomi in publishing training code and task environments alongside weights.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence58
- Adoption62
- Hype gap+8
- Incentives30
- Confidence60
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
Peter Farkas, co-creator of the open source MongoDB alternative FerretDB and now CEO of Percona, spoke at Open Source Summit Europe in Prague on Wednesday and asked: "Don't use 'open weight' and 'open source' interchangeably."
- [2]
In August, open-weight models accounted for 56% of tokens processed through Vercel's AI Gateway.
- [3]
In August, open-weight models accounted for 60% of US-originating token consumption on OpenRouter, with Chinese-developed models accounting for the majority.
- [4]
An AI model's weights are the numerical parameters produced during training, encoding the patterns the model has learned.
- [5]
Making weights available means developers can download and run a model on their own infrastructure, even when its creator has not released the ingredients or process used to produce it.
- [6]
"You don't have the source code, you don't have the training data, you only have the output of these two. So why would we call this open source in the first place?"
- [7]
James Landay, director at Stanford's Institute for Human-Centered AI, argued in an article published by HAI in August that access to model weights alone falls short of what developers and researchers need from genuinely open source AI.
- [8]
"Open weights answer 'can I run this?' Open source answers 'can I trust this, improve it, and build the next thing on top of it?'"
- [9]
Landay argues the major AI labs are largely answering the first question (can I run this) while falling well short of the second.
- [10]
"Are open weights a bad thing? No, open weights are great. You can run your models in your own environment, you can experiment with them, and if you understand the risks, you can also use it in production. The problem is when open weights are positioned as, 'hey, this is as good as open source'."
- [11]
Xiaomi launched its MiMo-V2.6 models, livestreaming nearly a week of reinforcement-learning training through a public dashboard, and released more than 7,000 reinforcement-learning task environments along with its RL training code and technical documentation.
- [12]
Xiaomi's release offers more transparency into how the models were fine-tuned; whether that is enough for MiMo-V2.6 to qualify as open source AI is a separate question; the release illustrates that open AI releases can expose vastly different amounts of what went into building a model.
- [13]
Farkas points to DeepSeek as an example of the confusion: its models are widely described as open source although the company released model weights rather than everything required to reproduce the models from scratch.
- [14]
Farkas argues that without a commonly understood boundary, "open washing wins," and worries that accepting a looser meaning of open source for AI could weaken the definition for software more broadly.
- [15]
"Companies are going to get away with calling something Apache 2.0 that [for example] you can't use in the European Union."
- [16]
The Open Source Initiative published its first Open Source AI Definition in 2024, establishing criteria around the freedoms to use and other freedoms.
- [17]
"'Open source' is only going to remain 'open source' as long as we preserve the actual meaning and the freedoms that are behind open source, and open weights are not providing that."
- [18]
The AI industry has embraced the language of open source even as some of its most prominent open models offer developers only a portion of what that term has traditionally promised.
- [19]
Open-weight models carried more than half of tokens on both gateways cited in August.
Sources
1 independent publisher whose own reporting we read for this story.
Topics and entities
Follow any of these and your For You feed starts watching them — no settings page required.
Topics
- Open-Weight Model PublishingFollow
- Open washingFollow
- What Counts as Open-Source AIFollow
Entities
- Peter FarkasFollow
- PerconaFollow
- FerretDBFollow
- Vercel AI GatewayFollow
- OpenRouterFollow
- James LandayFollow
- Stanford HAIFollow
- XiaomiFollow
- MiMo-V2.6Follow
- DeepSeekFollow
- Open Source InitiativeFollow
- Open Source AI DefinitionFollow
- Duane O'BrienFollow
- Open Source Summit EuropeFollow