Recursive

Recursive self-improving superintelligence to automate knowledge discovery.

[Recursive @Recursive_SI]

Recursive Flint is here: our first model, built with the Eureka Machine. Fast, local coding. Open weights available on Hugging Face.

[Richard Socher @RichardSocher]

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.

[Richard Socher @RichardSocher]

"I used Flint to build a bunch of programs and found it to be both fast and very capable. In my experience, models of this size can be annoyingly error-prone or slow compared to Claude or Codex, but Flint nailed my coding requests every time." — @willknight, @WIRED

https://t.co/EvXe313QfJ

[Recursive @Recursive_SI]

@Alec_Coughlin 🚀🚀🚀

[Recursive @Recursive_SI]

@Amank1412 @kaif9998 🏆

[Recursive @Recursive_SI]

@stanfordnlp @RichardSocher @timshi_ai @ChengleiSi Thank you, @stanfordnlp! An amazing group right there 🙌

[Recursive @Recursive_SI]

"I used Flint to build a bunch of programs and found it to be both fast and very capable. In my experience, models of this size can be annoyingly error-prone or slow compared to Claude or Codex, but Flint nailed my coding requests every time," writes @willknight.

[WIRED @WIRED]

Recursive makes the case that the benefits of using AI to build AI outweigh the risks. https://t.co/wgWQJrWZSW

[Recursive @Recursive_SI]

@Thom_Wolf 🤗

[NVIDIA RTX Spark @NVIDIARTXSpark]

Congrats @Recursive_SI on launching Flint 40B MoE trained on NVIDIA GB300 NVL72! 🙌

The open weight model was built with Eureka Machine, their automated AI research system that proposes ideas, tests them, and learns what to try next, all in under 4 months.

Run the model locally on RTX Spark, DGX Spark, and RTX PRO 6000 Blackwell, with the NVFP4 checkpoint on @huggingface: https://t.co/brURociBX1

[Recursive @Recursive_SI]

@GVteam @RichardSocher Thank you, @GVteam! This is just the beginning.

[Recursive @Recursive_SI]

@boardyai @RichardSocher 👀 Stay tuned.

[Recursive @Recursive_SI]

@rafikiidot @robinstetic Thank you. And this is just the start 🔥

[Recursive @Recursive_SI]

@_rockt @nvidia Let's go! 🚀

[Recursive @Recursive_SI]

@jeffclune The best team! 🫶

[Recursive @Recursive_SI]

@RichardSocher This team ships! 🔥🔥🔥

And we're hiring.

[Josh Tobin @josh_tobin_]

A fun small win for automated research:

We found and helped fix some edge cases that could affect inference performance in vLLM and SGLang.

In our last blog post, we described a reward hacking judge we developed for performance optimization tasks.

While applying the judge to some new autoresearch work, it discovered that some FlashInfer (a library underpinning vLLM and SGLang) kernels used a hard coded value of -50,000 as a masked-attention sentinel -- even though valid QK values can be smaller.

Corner cases like this that are numerically wrong but silent can cause a huge amount of headache to find and fix, e.g., the historical debate around flash attention (https://t.co/WtoQH97rYM).

Great use case for AI.
https://t.co/aiTEQn8hR3

[Recursive @Recursive_SI]

https://t.co/IlLyWptydX

[Kanishka Narayan MP @KanishkaNarayan]

It was such a privilege to join cofounders, @_rockt and Alexey Dosovitskiy, and a phenomenal team to celebrate @Recursive_SI’s launch and $650m raise last night.

I have been struck by Tim’s exceptional ambition, experience and humility in the last few months. I am delighted he and team will be pushing the frontier of AI right here in the UK 🇬🇧

[Recursive @Recursive_SI]

https://t.co/IPHqzO5gWl