显示与弗林特的AI驱动研究过程的第一个结果

显示与弗林特的AI驱动研究过程的第一个结果

关键点

- AI启动Recursive有发布 Flint，是一种编程任务的语言模型，主要由内部AI系统开发，名为"Eureka Machine".

- 弗林特使用具有400亿参数(25亿活性)的混合专家架构，并在Apple和Nvidia硬件上为快速本地执行进行优化。

- 该型号在Apache 2.0许可证下可在Hugging Face上免费获得。 Recurive的长期目标是利用这种方法来创建能够掌管自身进一步发展的AI系统。

AI startup Recursive用一个AI系统构建了它的第一语言模型，该系统提出，运行，并评价自己的实验。

AI startup Recursive拥有其第一语言模型发布 Flint. 一个叫做"Eureka Machine"的AI系统参与了几乎每一个发展步骤，自己提出想法，建立，测试，并根据结果规划后续实验。

弗林特(Flint)是一种专家混合模型(MoE)，旨在在当地硬件上进行代理编码，建造时间为四个月。 在400亿总参数中，每个符号只有25亿有效。 递归型声称 Flint引导编码型基准测试，如SWE-Bench Pro v2,Terminal-Bench 2.1，以及SWE-Bench 验证型在MOE模型中(如Qwen3.6-35B-A3B和Laguna S 2.1)，参数不足500亿。 结果还没有被独立核实。

该模型意在与250至400亿参数范围内的密集模型竞争，同时运行速度超过5倍。 在使用M4 Pro的MacBook Pro上，弗林特根据Recursive在短背景下每秒生成超过125个令牌。 在Nvidia RTX Spark与64GB的内存上，它是超过 150. 与类似模型相比，推论速度高达两倍。

Eureka Machine 自行选择了架构、量化和计算内核

Eureka机器将整个研究周期自动化，平行运行多种方法。 对于弗林特，它帮助培训数据，优化器，建筑设计，和推论调试。 三个例子显示了系统如何塑造模型的性能。

从拥有350亿总参数和30亿有效参数的架构开始，Eureka Machine测试了不同的专家配置。 最终设计了400亿总参数和25亿活性参数，结果是在它测试的所有变体中，速度和编码性能之间的最佳权衡。

为了量化，系统违背了惯例，将输出层也减少到4位。 附加的蒸馏步骤是为了恢复失去的精度，根据Recursive，它需要的训练步骤大约是前一种方法的一半。 完成的NVFP4型号仅坐落在横跨基准测试的16位版本后面0.7分。

Eureka机器还为Apple硅和Nvidia RTX Spark建造了专门的计算内核。 在Apple的M4和M5芯片上，一个金属巨核在代号生成期间将整个前传合并为单一操作。 递归称这些跨越多个开发阶段的优化使得高推论速度成为可能。

弗林特在当地运行，船用开放权重

弗林特支持上下文windows最多512 000个令牌，尽管Recursive建议将其封顶为256 000个。 在Apache 2.0许可下，模式权重可在Hugging Face上获得。 支持的运行时间包括llama.cpp,Ollama,LM Studio,vLLM，以及SGLang. 递归还原发布 reaction，它自己的推论引擎用于Apple硅。

所有这一切都意味着模型可以在当地运行，而不向外部服务器发送代码。 递归性确实承认弗林特在复杂任务上比大型专有编码模型需要更多的指导。 公司建议将更大的任务分成更小的步骤。

递归就是打赌它能造出更好的AI

由理查德·索彻等人，包括前DeepMind研究者Tim Rocktäschel创立的创业者，于2026年5月以46.5亿美元估值的融资6.5亿美元出道。 当时，Recursive没有具体的技术成果可以显示。 Flint是公司第一个从其自动化研究过程中公开使用的模型。

长期目标是建立不断自我改善的超智能。 每个新模型都应该使Eureka机器成为更好的研究者，因此下一代模型可以更快更高效地建造。 最终，Recursive希望将同样的方法应用于其他科学领域的自动化研究。

这实际上是否会产生加速的自我改善还远未解决。 前Google DeepMind研究带领Oriol Vinyals认为AI协助的自我改进是现实的，但认为智能爆炸不太可能。 对人工智能系统来说，寻找良好的研究想法和可靠判断结果仍然是特别困难的问题。

其他AI实验室也已经在使用AI来构建新的模型。 Anthropic 称 Claude现在处理自己开发工作的一大部分，而OpenAI则在其整个研究管道上部署AI agents. 这个自动化到底走了多远还有待辩论。 一项对700多个任务日志的研究发现AI agents处理大部份的作品，但人类仍然能做出大多数关键的研究呼声。

这些步骤是Recursive想与Eureka机器自动化。 弗林特是第一个具体的结果，但它还没有证明完全的递归自我改进。 为此，公司需要表明，用Eureka机器建造的模型实际上使Eureka机器本身成为更好的研究人员。

AI News Without the Hype – 由人类治愈

订阅THE DECODER免费阅读，每周AI通讯，我们独家的"AI雷达"前沿报告每年6次，全面存档访问，访问我们的评论部分。

现在订阅

本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。

Recursive shows first result of its AI-driven research process with Flint

Key Points

- The AI startup Recursive has released Flint, a language model for programming tasks that was developed largely by an internal AI system called "Eureka Machine."

- Flint uses a mixture-of-experts architecture with 40 billion parameters (2.5 billion active) and is optimized for fast local execution on Apple and Nvidia hardware.

- The model is freely available on Hugging Face under the Apache 2.0 license. Recursive's long-term goal is to use this approach to create AI systems that can take charge of their own further development.

AI startup Recursive built its first language model with an AI system that proposes, runs, and evaluates its own experiments.

AI startup Recursive has released Flint, its first language model. An AI system called "Eureka Machine" was involved in nearly every step of development, proposing ideas on its own, building them out, testing them, and planning follow-up experiments based on the results.

Flint is a mixture-of-experts model (MoE) designed for agentic coding on local hardware, and it was built in four months. Out of 40 billion total parameters, only 2.5 billion are active per token. Recursive claims Flint leads coding benchmarks like SWE-Bench Pro v2, Terminal-Bench 2.1, and SWE-Bench Verified among MoE models (such as Qwen3.6-35B-A3B and Laguna S 2.1) with fewer than 50 billion parameters. The results haven't been independently verified yet.

The model is meant to compete with dense models in the 25 to 40 billion parameter range while running more than five times faster. On a MacBook Pro with M4 Pro, Flint generates more than 125 tokens per second at short contexts, according to Recursive. On an Nvidia RTX Spark with 64 GB of memory, it tops 150. Compared to similar models, inference is up to twice as fast.

Eureka Machine chose the architecture, quantization, and compute kernels on its own

Recursive says the Eureka Machine automates the entire research cycle, running multiple approaches in parallel. For Flint, it helped with training data, optimizers, architecture design, and inference tuning. Three examples show how the system shaped the model's performance.

Starting from an architecture with 35 billion total and 3 billion active parameters, the Eureka Machine tested different expert configurations. The final design of 40 billion total and 2.5 billion active parameters turned out to be the best tradeoff between speed and coding performance among all the variants it tested.

For quantization, the system went against common practice and reduced the output layer to four bits as well. An added distillation step was meant to recover the lost accuracy, and according to Recursive, it needed roughly half as many training steps as the previous method. The finished NVFP4 model sits only 0.7 points behind the 16-bit version across benchmarks.

The Eureka Machine also built specialized compute kernels for Apple Silicon and Nvidia RTX Spark. On Apple's M4 and M5 chips, a Metal megakernel combines the entire forward pass during token generation into a single operation. Recursive says these optimizations across multiple development stages are what enable the high inference speed.

Flint runs locally and ships with open weights

Flint supports context windows of up to 512,000 tokens, though Recursive recommends capping it at 256,000. The model weights are available on Hugging Face under the Apache 2.0 license. Supported runtimes include llama.cpp, Ollama, LM Studio, vLLM, and SGLang. Recursive also released rec, its own inference engine for Apple Silicon.

All of that means the model can run locally without sending code to external servers. Recursive does acknowledge that Flint needs more guidance on complex tasks than large proprietary coding models. The company recommends breaking bigger tasks into smaller steps.

Recursive is betting it can build AI that keeps building better AI

The startup, founded by Richard Socher and others including former DeepMind researcher Tim Rocktäschel, came out of stealth in May 2026 and raised $650 million at a $4.65 billion valuation. At the time, Recursive had no concrete technical results to show. Flint is the company's first publicly available model from its automated research process.

The long-term goal is to build a superintelligence that keeps improving itself. Each new model is supposed to make the Eureka Machine a better researcher, so the next generation of models can be built faster and more efficiently. Eventually, Recursive wants to apply the same approach to automate research in other scientific fields.

Whether this will actually produce accelerating self-improvement is far from settled. Former Google DeepMind research lead Oriol Vinyals considers AI-assisted self-improvement realistic but thinks an intelligence explosion is unlikely. Finding good research ideas and reliably judging results remain especially hard problems for AI systems.

Other AI labs are already using AI to build new models too. Anthropic says Claude now handles a large share of its own development work, while OpenAI deploys AI agents across its entire research pipeline. How far that automation actually goes is still debated. A study of more than 700 task logs found that AI agents handle large portions of the work, but humans still make most of the critical research calls.

Those are the steps Recursive wants to automate with the Eureka Machine. Flint is a first concrete result, but it doesn't yet prove full recursive self-improvement. For that, the company would need to show that a model built with the Eureka Machine actually makes the Eureka Machine itself a better researcher.

AI News Without the Hype – Curated by Humans

Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive "AI Radar" frontier report six times a year, full archive access, and access to our comment section.

Subscribe now

Auto-fetched and translated by AI Atlas. All rights belong to the original author.