位于雷德伍德城的 AI 实验室 Inception 于周一发布了 Mercury 3。这是该公司所谓“扩散 LLM”的最新版本，其设计有别于大多数聊天机器人背后的系统。
Inception, the Redwood City AI lab, released Mercury 3 on Monday. It is the newest version of what the company calls a diffusion LLM, a design that differs from the systems behind most chatbots.
大多数语言模型属于自回归模型。它们一次生成一个 token，每个 token 的选择都取决于前一个。扩散模型的工作方式更接近同名的图像生成器。它从粗糙、含噪的输出开始，并行地同时优化许多 token，直到文本变得清晰。Inception 表示，正是这种并行方法让其模型运行得更快。
Most language models are autoregressive. They write one token at a time, each one chosen after the last. A diffusion model works more like the image generators that share the name. It starts with rough, noisy output and refines many tokens at once, in parallel, until the text resolves. Inception says that parallel approach is why its model runs faster.
根据公司自行基准测试，Mercury 3 在一系列测试中击败了 OpenAI 的 GPT-6 Luna、Anthropic 的 Claude Haiku 5.5 以及智谱的 GLM-5.3 Flash，同时运行速度快四到八倍。Inception 称，在 Nvidia 芯片上，其生成速度超过每秒 1,000 个 token。这些数据为公司自行报告。目前尚无外部机构发布过对该模型的测试。
On the company’s own benchmarks, Mercury 3 beats OpenAI’s GPT-6 Luna, Anthropic’s Claude Haiku 5.5 and Zhipu’s GLM-5.3 Flash across a range of tests, while running four to eight times faster. Inception says it generates more than 1,000 tokens a second on Nvidia chips. Those figures are self-reported. No outside group has published a test of the model yet.
该实验室表示，Mercury 3 的规模是其首个模型的 40 多倍，在标准基准测试中的能力约为首个模型的三倍，且速度保持不变。其上下文窗口（即一次能处理的文本量）已扩大至 100 万 token，高于上一版本的 26 万 token。
The lab says Mercury 3 is more than 40 times the size of its first model and about three times as capable across standard benchmarks, at the same speed. Its context window, the amount of text it can handle at once, has grown to one million tokens, up from 260,000 in the previous version.
速度优先，价格次之 Mercury 3 的定价为每百万输入 token 0.20 美元，每百万输出 token 0.70 美元。为配合发布，公司将两项价格均减半，降至 0.10 美元和 0.35 美元。Inception 未说明该促销将持续多久。
Speed first, price second Mercury 3 costs $0.20 per million input tokens and $0.70 per million output tokens. For launch, the company is halving both, to $0.10 and $0.35. Inception has not said how long that promotion will last.
主打的是速度，而非价格。回答更快的模型更易于大规模运行，这对需要连续多次调用模型的智能体和编码工具最为重要。
Speed, not price, is the pitch. A model that answers faster is easier to run at scale, which matters most for agents and coding tools that call a model many times in a row.
“它与 Luna 和 Haiku 模型相当，但速度快得多，”Inception 首席执行官 Stefano Ermon 对《The Deep View》表示。“它能达到每秒 1,000 个 token。从这些模型获得答案的方式几乎是即时的。”发布价格在输出 token 上低于竞品，并延续了整个行业更广泛的一轮降价。
“It’s comparable to the Luna and Haiku models, but it’s way faster,” Inception chief executive Stefano Ermon told The Deep View. “It’s getting 1,000 tokens per second. It’s almost instant in the way you get answers from these models.”The launch price undercuts those rivals on output tokens, and extends a wider run of price cuts across the industry.
Inception 还介绍了该模型的构建过程。Ermon 告诉 The Deep View，公司利用 Mercury 3 智能体群来改进模型本身，这是一种递归式自我改进。他表示，该技术在团队手工优化取得的成果之上，又使模型速度提升了约 16%。据该媒体报道，Inception 拒绝透露模型的总参数量。
Built with its own agents Inception also described how it built the model. Ermon told The Deep View that the company used swarms of Mercury 3 agents to improve the model itself, a form of recursive self-improvement. He said the technique added about 16% to the model’s speed, on top of the gains his team made by hand. Inception would not say how many parameters the model has in total, the outlet reported.
扩散模型仍属少数派押注。Google DeepMind 于 6 月发布了扩散文本模型 DiffusionGemma，表明大型实验室也在测试同类思路。该领域主流仍为自回归模型。独立测试将显示速度宣称在 Inception 自有基准之外是否站得住脚。
Diffusion is still a minority bet. Google DeepMind released a diffusion text model, DiffusionGemma, in June, a sign that larger labs are testing the same idea. Most of the field remains autoregressive. Independent testing will show whether the speed claims hold up outside Inception’s own benchmarks.
该模型现已通过 Inception 自有 API、OpenRouter、Baseten 以及 Vercel 的 AI Gateway 提供。它采用与 OpenAI 相同的 API 格式，因此开发者只需极少修改即可切换现有代码。Inception 的投资方包括 Menlo Ventures、微软 M12、英伟达风投部门，以及投资人 Andrew Ng、Andrej Karpathy 和 Eric Schmidt。
The model is available now through Inception’s own API, plus OpenRouter, Baseten and Vercel’s AI Gateway. It uses the same API format as OpenAI’s, so developers can switch existing code with little change. Inception’s backers include Menlo Ventures, Microsoft’s M12, Nvidia’s venture arm and the investors Andrew Ng, Andrej Karpathy and Eric Schmidt.