Thinking Machines amps up its bet against one-size-fits-all AI with its first open model, Inkling
Thinking Machines amps up its bet against one-size-fits-all AI with its first open model, Inkling
Thinking Machines Lab, the AI startup founded by former OpenAI CTO Mira Murati, released its first in-house AI model Wednesday morning, called Inkling. And unlike the flagship models from OpenAI, Anthropic, or Google, it’s open-weight, meaning outside developers and companies can download it and modify it directly.
由前 OpenAI 首席技术官 Mira Murati 创立的 AI 初创公司 Thinking Machines Lab 于周三上午发布了其首款自研 AI 模型，名为 Inkling。与 OpenAI、Anthropic 或 Google 的旗舰模型不同，它采用开放权重（open-weight）模式，这意味着外部开发者和公司可以直接下载并对其进行修改。
Inkling is a mixture-of-experts system with 975 billion total parameters, though it only draws on a fraction of that — about 41 billion — for any given task, a common design that keeps very large models faster and cheaper to run. It was trained on 45 trillion tokens of text, image, audio, and video, and reasons natively across all four, according to the company’s own release materials. For now, though, its outputs are limited to text, including code, styled artifacts, and structured data.
Inkling 是一个混合专家（mixture-of-experts）系统，总参数量为 9750 亿，但在执行任何特定任务时仅调用其中的一小部分——约 410 亿。这是一种常见的设计，旨在让超大规模模型运行得更快、成本更低。根据该公司发布的材料，该模型基于 45 万亿个文本、图像、音频和视频 token 进行训练，并能在这四种模态间进行原生推理。不过目前，其输出仅限于文本，包括代码、格式化文档和结构化数据。
The model is Thinking Machines Labs’ first public proof point after a year and a half spent building AI infrastructure largely out of public view. Some of that work had already surfaced in a May research preview of “interaction models” — AI designed to listen and speak (and even interrupt) instead of stop and wait as with typical chatbots. It’s also a test of the central bet behind the startup, which is that AI that organizations can adapt for themselves will outperform the one-size-fits-all models the biggest labs currently sell.
该模型是 Thinking Machines Labs 在过去一年半里低调构建 AI 基础设施后的首个公开成果。其中部分工作已在五月份的“交互模型”研究预览中有所体现——这种 AI 旨在倾听和说话（甚至可以打断对话），而不是像传统聊天机器人那样“停顿等待”。这也是对该初创公司核心赌注的一次考验：即企业能够自行适配的 AI，将优于目前各大实验室所售卖的“一刀切”式模型。
Inkling is designed to give calibrated answers, including flagging uncertainty rather than guessing, and lets users dial “thinking effort” up or down when they want to trade for speed. On one benchmark, the company says, Inkling uses a third as many tokens as Nvidia’s Nemotron 3 Ultra — its latest generation open-weight model — to hit the same coding performance. Thinking Machines doesn’t claim Inkling is best-in-class. Its newest blog post states explicitly that Inkling is “not the strongest overall model available today, open or closed.” What it’s evidently going for instead is well-rounded performance.
Inkling 旨在提供经过校准的答案，包括在不确定时进行标记而非盲目猜测，并允许用户根据对速度的需求调高或调低“思考力度”。据该公司称，在某项基准测试中，Inkling 达到同样的编码性能所消耗的 token 仅为 Nvidia 最新一代开放权重模型 Nemotron 3 Ultra 的三分之一。Thinking Machines 并不声称 Inkling 是同类产品中最强的。其最新的博客文章明确指出，Inkling “并非当今市面上最强的模型，无论是开源还是闭源”。显然，它追求的是全面均衡的性能。
That raises the question of who, within the enterprise market it’s targeting, this product is really for. Thinking Machines is, for now, marketing Inkling less as a finished product than as a starting point, something for organizations to fine-tune themselves through Tinker, the company’s model-customization platform. This also means customers, not Thinking Machines, are responsible for making sure their customizations are safe, for example. (Fine-tuning requires serious machine-learning talent.)
这引发了一个问题：在其瞄准的企业市场中，这款产品究竟是为谁准备的？目前，Thinking Machines 将 Inkling 更多地定位为一个起点，而非成品，旨在让组织通过该公司的模型定制平台 Tinker 进行自行微调。这也意味着，例如确保定制内容的安全性，责任在于客户而非 Thinking Machines。（微调需要具备深厚的机器学习人才储备。）
OpenAI, Anthropic, and Google have all taken a very different approach with ChatGPT, Claude, and Gemini, respectively, which were all built to compete as general-purpose chatbots first, with agentic, autonomous features layered on top. A post published by Thinking Machines last week was clearly meant as the backdrop for this release. AI that’s trained centrally by one company and then set in stone, the company argued in that post, underperforms AI that organizations shape themselves because so much expertise is specific to the people who hold it.
OpenAI、Anthropic 和 Google 在 ChatGPT、Claude 和 Gemini 上采取了截然不同的策略，它们首先被构建为通用聊天机器人，并在其之上叠加代理（agentic）和自主功能。Thinking Machines 上周发布的一篇文章显然是此次发布的背景铺垫。该公司在文中指出，由一家公司集中训练并“定型”的 AI，其表现不如组织自行塑造的 AI，因为大量的专业知识是特定于掌握这些知识的人员的。
Other arguments against closed models are gaining steam. In a blog post published Sunday, Microsoft CEO Satya Nadella — whose company has invested billions in both OpenAI and Anthropic — warned that enterprises using proprietary AI models effectively pay twice: once in subscription costs, and again by handing over business knowledge embedded in their prompts and corrections, which can be absorbed into future model versions. Hugging Face CEO Clem Delangue made a similar prediction in conversation with TechCrunch last week.
反对闭源模型的其他论点也正获得支持。微软首席执行官 Satya Nadella（其公司已向 OpenAI 和 Anthropic 投资数十亿美元）在周日发布的博客文章中警告称，使用专有 AI 模型的企业实际上支付了双重成本：一是订阅费用，二是通过提示词和修正意见交出了嵌入其中的商业知识，而这些知识可能会被吸收进未来的模型版本中。Hugging Face 首席执行官 Clem Delangue 上周在与 TechCrunch 的对话中也做出了类似的预测。
Frontier models, he said, will increasingly be reserved for experimentation and high-value tasks, while most production AI work shifts to private or open-source alternatives — the exact split Thinking Machines is building around. The clearest argument for Thinking Machines’ approach came from a recent project with Bridgewater Associates, the world’s largest hedge fund (which is not, for what it’s worth, a Thinking Machines investor). Researchers from both companies took an existing open-source model and trained it further on Bridgewater’s own financial expertise.
他表示，前沿模型将越来越多地被保留用于实验和高价值任务，而大多数生产环境的 AI 工作将转向私有或开源替代方案——这正是 Thinking Machines 围绕其构建的细分市场。支持 Thinking Machines 路径的最有力论据来自其与全球最大对冲基金桥水基金（Bridgewater Associates，顺便提一下，它并非 Thinking Machines 的投资者）最近的一个合作项目。双方的研究人员采用了一个现有的开源模型，并利用桥水基金自身的金融专业知识对其进行了进一步训练。
The result was said to score 84.7% on financial reasoning tests, beating top proprietary AI models, while costing roughly a fourteenth as much to run — though those results come from the two companies’ own evaluation, not an independent one. Either way, Thinking Machines is emphasizing how quickly it got here. OpenAI took roughly five years to bring its tech to market and show revenue, and Anthropic roughly three. Thinking Machines says it did the same in about nine months.
据称，该模型在金融推理测试中得分 84.7%，击败了顶尖的专有 AI 模型，而运行成本仅为后者的十四分之一左右——尽管这些结果来自两家公司自己的评估，而非独立评估。无论如何，Thinking Machines 都在强调其取得进展的速度。OpenAI 大约花了五年时间才将其技术推向市场并实现营收，Anthropic 大约花了三年。而 Thinking Machines 表示，它在九个月内就完成了同样的事情。
Some will wonder whether Inkling was trained on outputs from competitors’ models, a practice known as “distillation” that has drawn scrutiny across the industry. The short answer, per the company’s own materials, is partly. Thinking Machines pre-trained Inkling from scratch, but it says it used other open-weight models — including Moonshot AI’s Kimi K2.5 — to help generate some of its early post-training data before large-scale reinforcement learning took over. The next model, the company insists, will use fully self-contained post-training instead.
有些人会好奇 Inkling 是否使用了竞争对手模型的输出进行训练，这种被称为“蒸馏”（distillation）的做法在业内引起了广泛关注。根据该公司自己的材料，简短的回答是：部分使用了。Thinking Machines 从零开始预训练了 Inkling，但它表示在进行大规模强化学习之前，确实使用了其他开放权重模型（包括月之暗面 Moonshot AI 的 Kimi K2.5）来帮助生成部分早期训练后数据。该公司坚称，下一个模型将完全使用自包含的训练后数据。
On the cost side, Thinking Machines has been more guarded. It struck a partnership with Nvidia in March to deploy a gigawatt of Vera Rubin computing capacity and trained Inkling entirely on Nvidia’s GB300 NVL72 systems — but hasn’t said how it plans to cover those costs, and revenue, by most accounts, hasn’t been a priority. (A reported $50 billion fundraising round was said to be coming together last November but had stalled by January; the company has declined to talk about its funding picture since.) A related question is whether Thinking Machines’ spending will ever reach the scale of OpenAI’s or Anthropic’s, or whether its efficiency-driven approach means the economics look different. Put another way, the company’s bet may be less that it will eventually spend like its larger rivals than that it won’t need to at all — because once weig
在成本方面，Thinking Machines 则更为谨慎。它在三月份与 Nvidia 达成合作，部署了 1 吉瓦的 Vera Rubin 计算能力，并完全在 Nvidia 的 GB300 NVL72 系统上训练了 Inkling——但并未透露计划如何覆盖这些成本，且据多数报道称，营收并非其优先事项。（据报道，去年 11 月曾有一轮 500 亿美元的融资计划，但在 1 月份陷入停滞；此后该公司拒绝谈论其融资情况。）一个相关的问题是，Thinking Machines 的支出是否会达到 OpenAI 或 Anthropic 的规模，或者其以效率为导向的方法是否意味着经济模型有所不同。换句话说，该公司的赌注可能不在于最终会像其更大的竞争对手那样烧钱，而在于它根本不需要这样做——因为一旦……