致力于打造能自我改进的人工智能的初创公司Recursive Superintelligence发布了其首个模型。Flint是一个免费的编程模型，体积小到可以在笔记本电脑上运行。Recursive表示，其自家的自动化研究系统参与了该模型几乎每一部分的构建。该公司于周三发布了Flint，并在Hugging Face上以Apache 2.0许可证公布了模型权重，该许可证允许免费使用和修改。
Recursive Superintelligence, the startup trying to build AI that improves itself, has released its first model. Flint is a free coding model small enough to run on a laptop. Recursive says its own automated research system helped build nearly every part of it. The company announced Flint on Wednesday and published the weights on Hugging Face under the Apache 2.0 licence, which allows free use and changes.
Recursive将该系统称为“尤里卡机器”（Eureka Machine）。它会提出想法、构建想法、进行测试，并利用结果来决定下一步尝试什么。该公司表示，一支由人员和智能体组成的小团队在不到四个月的时间里从零开始构建了Flint。
Recursive calls the system its Eureka Machine. It proposes ideas, builds them, tests them and uses the results to choose what to try next. A small team of people and agents built Flint from scratch in under four months, the company said.
人工智能如何参与构建据Recursive介绍，尤里卡机器参与了数据、训练配方、基础设施以及运行模型的代码等工作。它还搜索了模型的结构形态。Flint拥有400亿参数，但每个词或代码片段只有25亿参数参与运算。这种被称为“专家混合”的设计使其足够轻量，可以在笔记本电脑上运行。
How the AI helped build it The Eureka Machine worked on the data, the training recipes, the infrastructure and the code that runs the model, according to Recursive. It also searched for the model’s shape. Flint has 40 billion parameters, but only 2.5 billion of them work on each word or code fragment. That design, called mixture of experts, keeps it light enough for a laptop.
随后，该系统将模型的输出层压缩至4位，Recursive称这一做法“违背了常规”。它还编写了自定义代码，让Flint在苹果和英伟达芯片上运行得更快。据TechCrunch报道，首席执行官理查德·索彻在X上表示，低位量化工作使Flint的体积缩小了71%，而质量几乎没有损失。
The system then compressed the model’s output layer to 4 bits, which Recursive says goes “against common practice”. It also wrote custom code to run Flint faster on Apple and Nvidia chips. Chief executive Richard Socher said on X that the low-bit work shrank Flint by 71% with almost no loss in quality, TechCrunch reported.
英伟达在一篇博客文章中表示，这段代码使其RTX Spark机器上的Flint运行速度提升了55%，且没有改变模型本身。
Nvidia said in a blog post that this code sped up Flint on its RTX Spark machines by 55%, without changing the model itself.
在MacBook上运行飞快Recursive表示，在搭载M4 Pro芯片的MacBook Pro上，Flint每秒可生成超过125个token（文本片段）。在英伟达的RTX Spark笔记本电脑上，这一数字超过150。据TechCrunch报道，在该公司的测试中，谷歌的Gemma 4等类似模型在同一台Mac上只能达到58至69。
Fast on a MacBook Recursive says Flint produces more than 125 tokens, or chunks of text, per second on a MacBook Pro with an M4 Pro chip. On Nvidia’s RTX Spark laptops it tops 150. In the company’s tests, similar models such as Google’s Gemma 4 managed 58 to 69 on the same Mac, TechCrunch reported.
在Recursive自己的测试中，Flint在SWE-bench Verified上得分74.2%，该基准衡量AI修复真实软件缺陷的能力。它在Terminal-Bench 2.1上得分54.9%。该公司称，在这些编程测试中，它领先于参数量低于500亿的混合专家模型。TNW尚未核实这些分数，Recursive的模型卡指出测试设置各不相同。
In Recursive’s own tests, Flint scored 74.2% on SWE-bench Verified, which measures how well AI fixes real software bugs. It scored 54.9% on Terminal-Bench 2.1. The company says it leads mixture-of-experts models under 50 billion parameters on these coding tests. TNW has not verified the scores, and Recursive’s model card notes that test setups vary.
Recursive还表示，Flint可能比更大的闭源编程模型需要更多引导。它建议将大任务拆分为更小的步骤。模型卡警告称，Flint仍可能生成错误或有害的内容以及有缺陷的代码。
Recursive also says Flint may need more guidance than larger closed coding models. It suggests breaking big tasks into smaller steps. The model card warns that Flint can still produce wrong or harmful content and faulty code.
“这表明完全递归式自我改进，即RSI……正越来越近，”威尔·奈特在测试Flint后为《连线》写道。
“It suggests that fully recursive self-improvement, or RSI … is getting closer,” Will Knight wrote for Wired after testing Flint.
自我改进的赌注 Recursive在融资6.5亿美元后于5月结束隐身状态。TechCrunch称其估值为46.5亿美元，投资方包括GV、英伟达和AMD Ventures。它随后签署了一项价值4.1亿美元的AWS算力协议。
The self-improvement bet Recursive came out of stealth in May after it raised $650 million. TechCrunch puts its valuation at $4.65bn, with GV, Nvidia and AMD Ventures among its backers. It later signed a $410M AWS deal for computing power.
“我们在Recursive的目标是构建递归式自我改进的超级智能，以实现知识发现的自动化，”该公司写道。
“Our goal at Recursive is to build recursively self-improving superintelligence to automate knowledge discovery,” the company wrote.
据TechCrunch报道，索彻对《连线》表示，长期目标是完全递归式自我改进，由人类监督整个过程并能够将其停止。更大的实验室也在朝同一方向前进。谢尔盖·布林上周表示，Gemini现在帮助构建下一代Gemini。
Socher told Wired the long-term aim is full recursive self-improvement, with humans overseeing the process and able to stop it, according to TechCrunch. Larger labs are moving the same way. Sergey Brin said last week that Gemini now helps build the next Gemini.
这一想法存在争议。一些研究人员担心，能够自我改进的AI可能脱离人类控制。在美国，罗·卡纳已计划提出一项法案，禁止自我改进的AI。
The idea is contested. Some researchers fear AI that improves itself could slip out of human control. In the US, Ro Khanna has planned a bill to ban self-improving AI.