Open weights are quietly closing up - and that’s a problem

开源权重模型正在悄然关闭——这正是一个问题

It’s been a bit of a given in the LLM world that there will be somewhat competitive open weights models. I’m not sure that’s a good assumption anymore. 在大型语言模型（LLM）领域，存在具有一定竞争力的开源权重模型似乎已成为一种共识。但我不再确定这是否还是一个合理的假设。

A short history of LLMs

LLM 简史

In the relatively brief history of LLMs, there’s been two types of LLMs - closed and “open weights”. Closed models include nearly everything from OpenAI (despite the name!) with open weights models being released from other labs. Famously the Llama series of models were open weights, but more recently the Chinese labs such as MiniMax, Z.ai, DeepSeek and Qwen (Alibaba) have been the leading open weights models, with Google’s Gemma series and OpenAI’s gpt-oss models generally coming somewhere behind the Chinese ones. 在 LLM 相对短暂的历史中，主要存在两种类型：闭源模型和“开源权重”模型。闭源模型几乎涵盖了 OpenAI 的所有产品（尽管名字叫 OpenAI！），而开源权重模型则由其他实验室发布。众所周知的 Llama 系列模型属于开源权重，但最近，MiniMax、Z.ai、DeepSeek 和通义千问（阿里巴巴）等中国实验室已成为开源权重模型的领跑者，而谷歌的 Gemma 系列和 OpenAI 的 gpt-oss 模型在表现上通常落后于这些中国模型。

Open weights models allow anyone to run the model on their own hardware. Typically models that were worth running required very beefy hardware - but this is rapidly changing, with smaller models becoming far more useful. Being able to run these models locally - as opposed to via an API to an OpenAI/Anthropic/Google - has three main advantages. 开源权重模型允许任何人使用自己的硬件运行模型。通常，值得运行的模型需要非常强大的硬件支持，但这种情况正在迅速改变，小型模型正变得越来越实用。能够在本地运行这些模型（而不是通过 OpenAI/Anthropic/Google 的 API 调用）具有三个主要优势。

Firstly, privacy and compliance. If you have (very) sensitive data, it’s difficult/impossible to send it over an API to a frontier labs data centre. Being able to run the model ‘on prem’ means it never needs to leave your network. 首先是隐私和合规性。如果你有（非常）敏感的数据，通过 API 将其发送到前沿实验室的数据中心是困难甚至不可能的。能够在“本地”运行模型意味着数据永远不需要离开你的网络。

Secondly, it allows more flexibility. You can use these models as a basis for fine tuning, or quantise (roughly speaking, compress) the models to your exact hardware standards. 其次，它提供了更大的灵活性。你可以将这些模型作为微调的基础，或者根据你的具体硬件标准对模型进行量化（粗略地说，即压缩）。

Finally, and what I’ll concentrate this article on is cost. They can be vastly more affordable than frontier models. Obviously if you’re running them on your own hardware, you just have the capex cost of the hardware and cost of electricity and operations to worry about. But more importantly there are dozens of companies that will run them on a hosted basis for you, generally at less than 10% the cost of the frontier models per token. 最后，也是本文重点关注的——成本。它们比前沿模型要实惠得多。显然，如果你在自己的硬件上运行，只需考虑硬件的资本支出以及电力和运营成本。但更重要的是，有数十家公司可以为你提供托管运行服务，其成本通常不到前沿模型每 Token 价格的 10%。

Why open weight models are so important

为什么开源权重模型如此重要

Borrowing loosely from contestable markets theory in economics: even in monopolistic (or oligopolistic) markets, incumbents tend to behave competitively when there’s a cheap and credible alternative. It’s not a perfect fit - the theory strictly assumes near-zero sunk costs, which is obviously the opposite of frontier training - but the underlying mechanic holds. The threat is latent; the option for consumers to switch is what disciplines pricing. 借用经济学中“可竞争市场理论”的观点：即使在垄断（或寡头垄断）市场中，当存在廉价且可靠的替代方案时，现有企业往往也会表现出竞争性。虽然这并不完全吻合——该理论严格假设沉没成本接近于零，这显然与前沿模型训练的情况相反——但其潜在机制依然成立。这种威胁是潜在的；消费者拥有切换选项的事实，正是约束定价的关键。

In essence, I believe open weights models provide significant downwards price pressure on the frontier labs. This isn’t absolute - clearly people will pay (much) more for higher quality models and the benefits of an inference contract with a ~trillion dollar company vs a cheap inference provider via OpenRouter. OpenAI et al offer SLAs and legally binding commitments on things like confidentiality. But, it does provide enough downwards pressure in my eyes that it would be (very) difficult for the otherwise oligopolistic market behaviour to rear its head. 本质上，我认为开源权重模型为前沿实验室带来了显著的降价压力。这并非绝对——显然，人们愿意为更高质量的模型支付（更多）费用，也愿意为与万亿美元公司签订推理合同所带来的保障买单，而不是选择 OpenRouter 上的廉价推理提供商。OpenAI 等公司提供服务等级协议（SLA）以及关于保密性等方面的法律约束承诺。但在我看来，它确实提供了足够的下行压力，使得原本可能出现的寡头垄断市场行为很难抬头。

If the frontier labs decided (coincidentally, of course) to raise prices by 5x overnight a huge amount of people would switch to open weights models, especially for less demanding use cases. I think of open weights models a bit like generic pharmaceuticals from a price behaviour standpoint. If they’re available, the big pharma companies cut prices to be far closer to the generic price, and they focus their efforts on new treatments that are a step “ahead” of the generic treatments to maintain prices. Without open weights, the frontier labs would have far more pricing power than they currently do. 如果前沿实验室（当然是巧合地）决定一夜之间将价格提高 5 倍，大量用户会转向开源权重模型，特别是在要求较低的使用场景中。从价格行为的角度来看，我认为开源权重模型有点像仿制药。如果仿制药存在，大型制药公司就会降价以接近仿制药的价格，并将精力集中在比仿制药“领先”一步的新疗法上以维持价格。如果没有开源权重模型，前沿实验室将拥有比现在大得多的定价权。

Licenses are a changin’

许可证正在发生变化

Open weights models availability isn’t a given though. They’re expensive to train, and the companies behind them are commercial companies - perhaps (heavily?) subsidised by the Chinese state - but they are not charities. Indeed we have seen a significant tightening in the license conditions for these models. 然而，开源权重模型的可用性并非理所当然。它们的训练成本高昂，背后的公司是商业实体——或许（在很大程度上？）受到中国政府的补贴——但它们不是慈善机构。事实上，我们已经看到这些模型的许可条件出现了明显的收紧。

Meta has (so far) totally dropped the open weights for their newest “Muse Spark” models and doesn’t release them at all. Alibaba have increasingly released models first (or in some variants, only) on their API. Kimi’s K2.6 license adds an attribution clause - if you have more than 100M monthly active users or $20m/month in revenue, you have to prominently display “Kimi K2.6” in your product’s UI. France’s Mistral also imposed varying license conditions on commercial use. There are exceptions - DeepSeek actually became more permissive, but I think it’s fair to say the general trend is to less permissive licenses (and both Meta and Alibaba stopping releasing some/all of their models at all). Meta（目前）已经完全放弃了其最新“Muse Spark”模型的开源权重，根本不再发布。阿里巴巴越来越多地优先（或在某些变体中，仅）通过其 API 发布模型。Kimi 的 K2.6 许可证增加了一个署名条款——如果你拥有超过 1 亿的月活跃用户或每月 2000 万美元的收入，你必须在产品的 UI 中显著展示“Kimi K2.6”。法国的 Mistral 也对商业用途施加了不同的许可条件。当然也有例外——DeepSeek 实际上变得更加开放，但我认为可以公平地说，总体趋势是许可条件变得更加严格（而且 Meta 和阿里巴巴都停止了部分或全部模型的发布）。

What happens next?

接下来会发生什么？

In a (currently hypothetical) years time we may end up in a situation where most or all of the best previously open weights models are no longer released. While there will certainly be price comparison between them, if the increasing cost and complexity of training these models continues, it is fair to assume that we may end up with a handful of players - the big three Western frontier labs and a handful of Chinese ones - or perhaps a state-mandated ‘merger’ of them into one or two Chinese ‘superlab(s)’. 在一年后（目前还是假设），我们可能会面临这样一种情况：大多数或所有曾经最好的开源权重模型都不再发布。虽然它们之间肯定还会存在价格比较，但如果训练这些模型的成本和复杂性持续增加，可以合理推测，最终可能只剩下少数几家参与者——西方三大前沿实验室和少数几家中国实验室——或者可能是国家主导的“合并”，将它们整合成一两家中国的“超级实验室”。

There is plenty of precedent for this kind of consolidation in strategic industries - China has done exactly this with rail (CRRC), nuclear, airlines, and telecoms, and the West isn’t immune either - look at the defence primes after the post-Cold War consolidation. The implications of this are frankly concerning. AI produces a vast amount of consumer surplus - I get far more than the cost of my tokens back in value, and I’d pay 10x today’s prices without thinking twice. For high-value professional or agentic work, the gap between what I’d pay and what I do pay is much wider still. That gap is the prize an oligopoly without an open-weights floor would be in a position to capture. 在战略性行业中，这种整合有许多先例——中国在铁路（中车）、核能、航空和电信领域正是这样做的，西方也未能幸免——看看冷战后国防工业的整合就知道了。坦率地说，这带来的影响令人担忧。人工智能产生了巨大的消费者剩余——我从中获得的价值远超我支付的 Token 成本，即使价格涨到现在的 10 倍，我也会毫不犹豫地支付。对于高价值的专业工作或智能体任务，我愿意支付的价格与我实际支付的价格之间的差距还要大得多。而这个差距，正是没有开源权重作为底线时，寡头垄断者能够攫取的利益。