Pinned Tweet
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 hs
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.
• 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 hs
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.
635
891
8,426
3,406,422
Richard Socher retweeted
Our felt sense of what it means to be human remains uniquely ours, and no machine will take that from us. The generation coming of age now understands this intuitively. They will see AI as a cognitive extension—a tool that amplifies their abilities rather than diminishes them.
~ @RichardSocher wrote a brilliant book, "The Eureka Machine". Everybody should read it.
~ @RichardSocher wrote a brilliant book, "The Eureka Machine". Everybody should read it.
0
1
5
1,160
Richard Socher retweeted
@RichardSocher @Recursive_SI This is a big step up over my local Qwen 3.8 setup—not just on code quality, but actually getting 125+ tok/s without babysitting the context window.
0
1
1
1,688
Richard Socher retweeted
I had a deeper look at Recursive's 40B Flint model huggingface.co/RecursiveSI/fl… I ran a few test yesterday but what I hadn't checked was the third-party support for the model. I was using
@Prince_Canuma
's mlx-vlm but at the time it wasn't doing justice to the model. I re-ran the merged PR 2569 (fused metal kernels), the different to stark, up to 30% which is amazing and knowing Prince, we'll see more. As an agentic model this is a beast, up there with the leaders and holding its own. Go team Recursive!
@Prince_Canuma
's mlx-vlm but at the time it wasn't doing justice to the model. I re-ran the merged PR 2569 (fused metal kernels), the different to stark, up to 30% which is amazing and knowing Prince, we'll see more. As an agentic model this is a beast, up there with the leaders and holding its own. Go team Recursive!
4
2
20
4,565
Richard Socher retweeted
It's hard working on product growth in 2026. Most of my end users are now AI agents, who are immune to well-worn tactics.
So when @getpostman made this new "AI-ready APIs Leaderboard" and put @youdotcom on the top, beating the legendary @OpenRouter by two points, I feel like we won the championship.
All the hard work over the past 13 months making our APIs, our platform, and the developer experience to world-class has really paid off.
Now, some recognition. Super proud to work at @youdotcom along side @edwardirby @itsakhilp @tyler5673 @king_chaga @srbh_shrm (who are a joy to work with and just show up everyday ready to build a great product.)
As @RichardSocher says "Better, better, never done" so we'll keep shipping. (Maybe improvements to our new router product... 😀)
postman.com/ai/ai-ready-ap…
So when @getpostman made this new "AI-ready APIs Leaderboard" and put @youdotcom on the top, beating the legendary @OpenRouter by two points, I feel like we won the championship.
All the hard work over the past 13 months making our APIs, our platform, and the developer experience to world-class has really paid off.
Now, some recognition. Super proud to work at @youdotcom along side @edwardirby @itsakhilp @tyler5673 @king_chaga @srbh_shrm (who are a joy to work with and just show up everyday ready to build a great product.)
As @RichardSocher says "Better, better, never done" so we'll keep shipping. (Maybe improvements to our new router product... 😀)
postman.com/ai/ai-ready-ap…
5
5
18
1,474
Richard Socher retweeted
Given the size, I was surprised to see that this does work reasonably well on my home 4090 with 24GB of RAM. With the nvp4 version and offloading some layers to cpu, it manages 50 tokens/s even after a 200k token prefill.
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 hs
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.
3
2
34
3,827
Richard Socher retweeted
@RichardSocher @Recursive_SI I have been working with Flint extensively over the past 24 hours via OpenRouter (I'm eager to try it on my Mac), and it is fantastic.
Great work!!
Great work!!
0
1
3
983
Richard Socher retweeted
we’re partnering with @youdotcom to bring Wafer’s inference to developers through its new Inference Router.
access Wafer through one OpenAI-compatible API, with multi-provider failover and built-in web search from You.com.
link in thread 🧵
access Wafer through one OpenAI-compatible API, with multi-provider failover and built-in web search from You.com.
link in thread 🧵
Introducing Inference Router: enterprise-grade inference on demand.
Kimi K3, GLM 5.3, DeepSeek V4.1 Flash, Qwen3.8 Flash, Muse Spark 1.3, Mercury 3 and more, available with an uptime SLA of at least 99.9%. US hosting and web search built in.
you.com/router
8
7
40
4,047
Richard Socher retweeted
Just tested this out and I'm stooked. So far it looks to be my preferred local model for now. We'll see how qwen4 compares when it comes out.
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 hs
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.
0
1
2
1,623
Richard Socher retweeted
I've been a heavy Qwen user. Now 100% on @Recursive_SI 's Flint. I built this side by side test, running both models at 8bit on the same hardware (RTX A6000 48Gb) try it: rotem.ai/flintvsqwen/ (This is running locally on two identical cards) BTW 4 bit is even faster.
OK, I've had a little play with this 40B MoE Flint model, here are my findings.
It's a good model but at 40B it needs to be compared to models a similar size and the Qwen3.6-35B MoE and Qwen3.8-27B are both smaller so make ideal candidates.
See thread below...
5
5
26
11,918
Introducing the @youdotcom Inference Router: enterprise-grade inference for leading LLMs, on demand.
• Minimum of 99.9% uptime SLA, backed by service credits
• Zero data retention across every model call and web search request
• 100% US-hosted infrastructure, with •Built-in web search
One OpenAI-compatible API, with no platform fee.
• Minimum of 99.9% uptime SLA, backed by service credits
• Zero data retention across every model call and web search request
• 100% US-hosted infrastructure, with •Built-in web search
One OpenAI-compatible API, with no platform fee.
6
6
88
6,609
Leading models from @deepseek_ai, @Kimi_Moonshot, @Zai_org, @Alibaba_Qwen, @MiniMax__AI, @AIatMeta, and more, including Kimi K3, GLM 5.3, and Qwen3.8 Flash.
Switching is one line of code: change your base URL, and your existing OpenAI-compatible code runs on Inference Router.
Switching is one line of code: change your base URL, and your existing OpenAI-compatible code runs on Inference Router.
1
1
2
921
Richard Socher retweeted
Whoops! Ik heb Flint geïnstalleerd op mijn MacMini M4 Pro 48GB, en het is echt snel, en de kwaliteit is vergelijkbaar met Qwen38! Werkt ook goed met Claude.
Voor de leken: ik heb nu mijn eigen gratis ChatGPT, en al mijn data blijft van mij!
(Nou ja gratis... je moet wel een dikke MacMini kopen nog.. maar het is een kwestie van tijd, een paar jaar nog dan heeft iedereen het in zijn telefoon zitten. Ik vraag me wel af hoe de grote jongens dan nog hun geld moeten verdienen...)
Voor de leken: ik heb nu mijn eigen gratis ChatGPT, en al mijn data blijft van mij!
(Nou ja gratis... je moet wel een dikke MacMini kopen nog.. maar het is een kwestie van tijd, een paar jaar nog dan heeft iedereen het in zijn telefoon zitten. Ik vraag me wel af hoe de grote jongens dan nog hun geld moeten verdienen...)
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 hs
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.
3
1
16
3,830
Richard Socher retweeted
Tested this on my M4 Max. Flint vs the same Qwen 27B locally: both went 6/6 on my agent tasks, Flint was 5x faster (33s vs 164s a task). Put Qwen on @cerebras and it's 3.5s. @RichardSocher @andrewdfeldman
Here we go! Purpose built AI generated LLMs!
Then AI helped optimize the code that runs it.
Recursive released Flint, a new coding/agent model designed specifically for fast local inference.
Beats Qwen3.5/3.6/3.8 in some coding benchmarks!
Stats …
🧠 39.8B total parameters
en
📚 262K documented context
🤖 reasoning + tool use
🔓 Apache 2.0
And there are official GGUFs:
🟢 NVIDIA NVFP4 → 22.9GB
🍎 Apple INT4 → 26.0GB
Recursive recommends 48GB-class memory for these local builds, so think:
🍎 48GB+ Apple Silicon
🟢 48GB+ NVIDIA
⚡ 64GB RTX Spark
Recursive reports its optimized runtime at
🍎 M4 Pro → >125 TPs
🟢 64GB RTX Spark → >150 TPs
And WIRED’s Will Knight actually ran Flint locally on a 48GB Mac and reported it was fast and successfully handled the coding jobs he tried.
Side Note: Recursive built Flint using its Eureka Machine, an automated AI research system that helped
🧩 design the architecture
🧪 develop training recipes
⚙️ optimize infrastructure
💻 write hardware-specific inference kernels
Nvidia says those specialized kernels improved Flint’s RTX Spark generation performance by 55% wo changing the model weights.
Recursive-reported coding results
SWE-bench Verified → 74.2
SWE-bench Pro v2 → 89.7
Terminal-Bench 2.1 → 54.9
LiveCodeBench v6 → 80.3
So this isn’t merely AI writing more AI code.
Link in ALT
1
1
4
2,240
Richard Socher retweeted
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 hs
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.
• 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 hs
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.
635
891
8,426
3,406,422
Richard Socher retweeted
Flint on a desk-sized GB10: 95 tok/s, still 37.5 at 128K context, and 9–10× faster than a dense 27B on the same machine. HumanEval 160/164.
Measured locally the day llama.cpp merged Flint support
@HP #ZGX @nvidia #GB10 @Recursive_SI
Measured locally the day llama.cpp merged Flint support
@HP #ZGX @nvidia #GB10 @Recursive_SI
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 hs
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.
6
1
13
1,994
Cool to see a bunch of project examples being done by our local model Flint.
Incredibly fun journey building the RSI loop(s) that led to a best-in-class local coding LLM.
A few simple coding examples Flint could do in < 10 mins.
0 endpoint. 0$. Full CoTs.
Stay tuned for more!
ps: Fable made this video from the artifacts alone. No editing.
2
0
59
6,101