Introducing Flint, Recursive’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 Hugging Face, run it locally on your favorite framework or try it for free on OpenCode or Pi.show more

Flint pushes the Pareto frontier on local coding.

Flint is the best-performing model in its class (MoE models under 50B) on agentic coding benchmarks like SWE-bench Pro v2, Terminal-Bench 2.1, and SWE-bench Verified. It also rivals similarly sized dense models that are much slower to run.

Flint is also the fastest model in its class: • 125+ tokens/sec on an M4 Pro MacBook Pro • 150+ tokens/sec on a 64GB NVIDIA RTX Spark Up to 2x faster than comparable models on short contexts, and 5x+ faster than dense models with similar coding scores. Your data stays on your computer, and it works offline.

Flint is also a real-world test of the Eureka Machine, our system for automated AI research. It runs the full research loop: ideation, implementation, and validation. Then it uses the results to decide what to try next. We used it in nearly every part of how Flint was built. This is what allowed our very small team to build this model in <4 months. Its contributions include a sparser model shape, low-bit training that shrank Flint by 71% with almost no loss in quality and custom kernels.

Try Flint today. Read the blog: Download on Hugging Face: Run Flint locally with @ollama, @lmstudio, @UnslothAI, @exolabs, @Nativ_AI, @sgl_project, @vllm_project and llama.cpp or try it for free on @opencode and @pidotdev

@Recursive_SI Huge congrats, @RichardSocher and team! 🔥🔥🔥

@Recursive_SI 125 tokens a second on an m4 pro💀

@Recursive_SI

@Recursive_SI Congratulations on Flint! I'm a big advocate for local AI and spend a lot of time using coding models for a variety of tasks on my quad RTX 6K Pro system. I'll definitely be giving this one a shot. Excited to see what you all cook up next!

@Recursive_SI congrats! Really tough to beat Qwen and already there

@Recursive_SI Good alternative for those who don't want to use (or can't use) closed models.

@Recursive_SI the future of vibe coding is local

@Recursive_SI How much does running the model locally change the privacy equation for coding agents?

@Recursive_SI 125 tokens/second on a macbook is crazy. wow

@Recursive_SI Fast coding models are useful, but fast models that can actually solve real coding problems change the equation.

@Recursive_SI Let's gooo

@Recursive_SI Love this. Fast local coding models that actually keep up with the benchmarks feel like the practical unlock a lot of people have been waiting for. Congrats on shipping it.

@josh_tobin_ @Recursive_SI Let’s go!! incredible on so many fronts: best in class model, eureka machine speed to frontier, unlock for local coding and agents. Congrats!

@Recursive_SI Building this in under four months is really impressive.

@Recursive_SI so what does the eureka machine build next

@Recursive_SI what's the minimum RAM you'd recommend to run Flint comfortably on a Mac?

@Recursive_SI congrats on shipping Flint

@Recursive_SI 🚀🚀🚀🚀

@Recursive_SI This could make local AI coding workflows much smoother.

@Recursive_SI Local coding at 125+ tokens/sec with open weights? This is exactly where AI gets exciting.

@Recursive_SI This reminds me a lot of meta-learning, but at the level of the entire research pipeline. I find this direction of automated research really interesting, especially if improvements can compound across successive generations.

@Recursive_SI Hum well done guys, the Eureka Machine part is what got me, btw which of its discoveries surprised the team the most while building Flint?

@Recursive_SI MASSIVE! Congrats :)

@Recursive_SI how much of the architecture search did the Eureka Machine run on its own vs the team steering it? curious where the human calls still mattered

@Recursive_SI Congrat🔥

@Recursive_SI Excited to see what developers build with Flint.

@Recursive_SI Local coding models are getting seriously impressive

@Recursive_SI A MacBook now runs near-27B-dense coding quality at 10x the speed. Recursive's own M4 Pro chart: Flint 131 tok/s vs Qwen3.8 27B at 13 (still 6.8x at 128K). SWE-bench Verified gap: 1.4 pts (74.2 vs 75.6). The catch is Terminal-Bench: 54.9 vs 73.0.

@Recursive_SI 125 tokens a sec on a macbook is impressive, congrats on the launch

@Recursive_SI On-device AI will be such a massive theme

@Recursive_SI Congrats, Richard, and to your team! Using the Eureka Machine to help build Flint itself is a great first test of automated research. Excited to run this locally and test it on some of our internal autoresearch evals.

@Recursive_SI Well done sir.

@Recursive_SI is the opencode free version same NVFP4 model or full 16 bit one

@Recursive_SI if AI can improve training and inference stack itself, future models could get much cheaper and faster to build...great work