Recursive releases its first model Flint: local coding MoE, 125+ tokens/s on a MacBook
English text is being prepared.
Recursive releases its first model Flint, a MoE with 40 billion total parameters and 2.5 billion activated parameters per token, aimed at local Agent coding, with weights open-sourced under Apache 2.0. With short context, it exceeds 125 tokens per second on an M4 Pro MacBook Pro and over 150 on a 64GB RTX Spark, and supports llama.cpp, Ollama, LM Studio, vLLM, SGLang, and Unsloth. Nearly every stage of its R&D involved the automated research system Eureka Machine, including architecture search, quantizing the output layer to 4 bit with offline distillation, and fusing Metal kernels for M4/M5.
Why read this
AI score 75
The original text explains Flint's local coding positioning and the three specific stages in which the Eureka Machine participated in R&D, showing how automated research translates into architecture, quantization, and kernels.
Topics
View original postIntroducing 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.
View original
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.
Body
Flint 是 @Recursive_SI 的第一个模型：总参数 400 亿、每个 Token 只激活 25 亿参数的 MoE，专门用于在本地电脑上做 Agent 编程，权重以 Apache 2.0 开源。它真正的看点不只是模型本身，更在于它几乎每个环节都由公司的自动化研究系统 Eureka Machine 参与完成，被当作"AI 自己做 AI 研发"的第一次实战检验。
模型信息 · 定位：在总参数低于 500 亿的 MoE 模型里，Agent 编程能力最强（SWE-Bench Pro v2、Terminal-Bench 2.1、SWE-Bench Verified）；编程分数和 250 亿到 400 亿参数的稠密模型相当，速度快 5 倍以上。 · 速度：短上下文下，M4 Pro MacBook Pro 上每秒超过 125 个 Token，64GB 的 RTX Spark 上每秒超过 150 个，最高是同类模型的 2 倍。 · 上下文：最长 512k，官方建议控制在 256k 以内。 生态：支持 llama.cpp、Ollama、LM Studio、vLLM、SGLang、Unsloth，另外推出了针对 M4 Pro 和 M5 优化的轻量推理引擎 rec。 · 官方给的使用建议：适合当日常的本地编程模型，数据不出本机、也不用付 API 费用。复杂任务要拆小、把指令写清楚，它比大型闭源模型更需要引导。
Eureka Machine 具体做了什么 它自动跑完整个研究循环：提出想法、实现、测试，再根据结果决定下一步，并行探索多个方向，把有效的思路组合起来。在 Flint 上有三个例子： 1. 架构搜索：从总参数 350 亿、激活 30 亿的基线出发，搜索专家数量和宽度的配置，最后选定总参数 400 亿、激活 25 亿，在速度和编程能力之间取得最佳平衡。 2. 反常规的量化：通常输出层（lm_head）不做量化，它把输出层也量化到 4 bit，原因是激活参数很少时，输出层占每个 Token 读取量的比例很大。为了补回精度，它又加了一步离线蒸馏：以 16 bit 模型为老师，让量化后的模型在自己采样的精选数据上训练，训练步数几乎减半。最终 NVFP4 版本和 16 bit 版本的分差在 0.7 分以内。 3. 内核优化：为低比特投影、MoE、注意力和预填充生成专用内核；在 M4、M5 上把整个解码前向过程融合成一个 Metal 大内核。