Recursive 发布 40B MoE 编码模型 Flint，MacBook Pro 上 125+ tokens/s
Automated AI research is now producing models people can download and run runs entirely on a local machine under an Apache 2.0 license.
Recursive 发布首个模型 Flint，这是一个 40B 混合专家编码模型，采用 Apache 2.0 许可，权重已在 Hugging Face 开放，由自动化研究系统 Eureka Machine 在不到 4 个月内从零构建。Flint 在 MacBook Pro 上以 125+ tokens/s 生成代码，在 RTX Spark 上达到 150+ tokens/s。在 48GB MacBook 上，它的平均质量约 81%，对比 Qwen3.8 27B 的约 85%，但在 128K token 长会话中速度约为后者的 7 倍。
推荐理由
精选AI 评分 76
这条内容给出了本地编码模型在速度与质量上的具体取舍数据，读者可据此判断本地部署的可行边界。
主题
查看原帖隆重介绍 Flint，@Recursive_SI 的首个模型：一个由我们的自动化研究系统 Eureka Machine 打造的快速本地编码模型。 • 同类最佳的快速本地编码 • 速度最高可达同类模型的 2 倍：在 MacBook Pro 上达到 125+ tokens/秒，在 RTX Spark 上达到 150+ tokens/秒 • 在不到 4 个月内从零构建 在你自己的机器上享受私密、无计量限制的智能。权重现已在 @huggingface 开放，可在你喜欢的框架上本地运行，或在 @opencode 或 @pidotdev 上免费试用。
查看原文
Introducing Flint, @Recursive_SI’s first model: a fast local coding model built with our automated research system, the Eureka Machine. • Best-in-class fast local coding • Up to 2x faster than comparable models: 125+ tokens/sec on a MacBook Pro, 150+ on RTX Spark • Built from scratch in under 4 months Private, unmetered intelligence on your own machine. Open weights available now on @huggingface, run it locally on your favorite framework or try it for free on @opencode or @pidotdev.
正文
Automated AI research is now producing models people can download and run runs entirely on a local machine under an Apache 2.0 license.
Recursive Flint, a 40B mixture-of-experts coding model, was built with the Eureka Machine, a loop that proposes ideas, implements them, tests them and uses the results to choose what to try next.
Flint writes code at 125+ tokens per second on a MacBook Pro, which works out to about 90 words a second, roughly 20x faster than you can read it.
And it has pushed the Pareto frontier on local coding.
On a 48GB MacBook, local models used to make you pick between smart but slow (Qwen3.8 27B scores highest but runs at about 10 tokens per second) and fast but weak (tiny Qwen3.5 models).
Recursive Flint sits almost as high on quality (about 81% average against about 85%) while running roughly 7x faster across a long 128K-token session, and the shaded area marks the speed and quality combination no model offered before.