Recursive open-sources 40B MoE coding model Flint, runnable on MacBook Pro

A 40B MoE "programmer" that you can take with you on a MacBook Pro!

After leaving Meta, Yuandong Tian is making a strong comeback—

With Recursive, the startup he founded, he has just open-sourced its first model, Flint:

Focused on local Agent coding, with 40 billion total parameters and only 2.5 billion activated per token generated.

And! It is also 2x faster than similar products.

Paired with an optimized inference engine, Flint can run at least 125 tokens/s on a MacBook Pro, and over 150 tokens/s on an RTX Spark.

This is also the first model deliverable from Recursive since its founding, produced with the help of an automated AI research system.

Even more impressive, the model was trained from scratch in just 4 months.

You have to hand it to Yuandong Tian and his team—they really know how to play the game.

Only this time, the first step has landed on developers' laptops.

It runs on a MacBook Pro

The first problem Flint has to solve is a very practical one for local models:

Fitting is not enough—it also has to run fast enough and get work done.

It uses a MoE mixture-of-experts architecture. Specifically, each MoE layer has 256 routed experts, 6 are selected per token, and there is also a shared expert that always works.

This way, the total parameter count is sufficient, while the activated parameters involved in each computation step can be steadily kept at 2.5 billion.

The model supports a context window of up to 512k tokens, and data can be stored locally for offline use.

At the same time, Recursive has also open-sourced the inference engine rec, specifically optimized for running Flint on Apple chips.

In official tests, using rec with the same Flint model improves token throughput by about 1.4–2.6x at corresponding context lengths compared with llama.cpp.

Its final speed is 5x faster than dense models of the same class (25–40 billion parameters), and even with short contexts it has nearly a 2x advantage.

This includes Muse Glimmer from his former employer Meta.

Its coding results are also impressive:

Flint scores 74.2% on SWE-bench Verified, 89.7% on SWE-bench Pro v2, and 54.9% on Terminal-Bench 2.1, all the best in its class.

Without exaggeration, Flint has broken through the Pareto frontier in local coding: fast inference speed and strong coding ability.

As for how to get it running, Recursive has also kept the setup process as short as possible.

After installing rec on a supported Mac, a single "rec launch pi" command automatically downloads the model, starts the service, and connects to coding tools; OpenCode also has a corresponding entry point.

Currently rec supports M4 Pro, M5 Pro, and M5 Max, requires at least 48GB of unified memory, and about 30GB of free disk space.

In other words, for developers who want to move some of their daily coding to local machines, Flint is indeed a good option.

Back to Recursive itself: the company only made its public debut in May.

As soon as it appeared, it secured $650 million (about 4.4 billion yuan) in early-stage funding, at a valuation of $4.65 billion (about 31.6 billion yuan).

Google GV and Greycroft co-led the round, with giants such as Nvidia and AMD also participating.

One bet is on the people.

All eight co-founders have impressive backgrounds. Besides Yuandong Tian, there are Richard Socher, a student of Andrew Ng; Tianlin Shi from Tsinghua's Yao Class; and Alexey Dosovitskiy, one of the authors of Vision Transformer.

The second bet is on the approach.

From the very beginning, Recursive made its research direction clear: RSI, or recursive self-improvement.

Eureka Machine is Recursive's initial attempt at this problem.

The system first proposes ideas for a goal, then implements the ideas, and decides the next step by running experiments and checking results. Multiple research paths can proceed simultaneously, and previously effective experience can be retained and combined.

The first batch of experimental results Recursive previously made public came from this.

On NanoChat Autoresearch, NanoGPT Speedrun, and Nvidia's SOL-ExecBench, it achieved SOTA in the overall and individual tests, respectively.

These three benchmarks also correspond in turn to training quality, training speed, and GPU kernel efficiency.

By the time of Flint, the problem became bigger.

The team combined improvements from these different stages into a model that could ultimately be delivered.

This also explains why Recursive chose local coding as the direction for its first model: local code tasks readily yield feedback, and the gains found by automated research are easier to check item by item.

Clearly, progress is good.

So for Yuandong Tian and Recursive, the next step is clear—

Let's wait and see.

Reference links:

This article comes from the WeChat public account "QbitAI", author: Luyu, published by 36Kr with authorization.