Teaching AI the joy of invention at @Recursive_SI, Professor of AI @AI_UCL, PI @BOLD_Lab_AI, Fellow @ELLISforEurope. Ex @GoogleDeepMind @AIatMeta @CompSciOxford
London, England
Joined December 2010
- Tweets6.1K
- Following1.9K
- Followers50.1K
- Likes20.8K
Pinned Tweet
My @iclr_conf 2025 Keynote on "Open-Endedness, World Models, and the Automation of Innovation" is now publicly available: iclr.cc/virtual/2025/invited…
Tim Rocktäschel retweeted
Just ran this on my Mac mini against my daily driver (Qwen 3.6 35B A3B) and I think we have a winner.
Tim Rocktäschel retweeted
1/ Can a model that runs on a laptop also drive GPU training research?
Recursive Flint, our new 40B open model, improved NanoChat across 700+ experiments—finishing close to our historical Claude Haiku 4.5 results.
We were pleasantly surprised. Here's what it tried 🧵
Tim Rocktäschel retweeted
Had my agent race Flint against Qwen3.8 27B on my Mac Studio with a couple real coding tasks.
Flint did 110 tok/s vs 28 and nailed every one. Qwen thought for 7 minutes on one and never answered.
And Recursive's fast kernels don't even support my M4 Max yet. New local default.
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.
Tim Rocktäschel retweeted
darkmux supports Flint and rec.
This table compares Flint on the Macbook Pro 128 GB M5 Max, with qwen 3.6 35B-A3B turboquant served by LM Studio.
The eval in this case is my very own pepper-grinder fixture. congrats to the @Recursive_SI on a successful launch.
Tim Rocktäschel retweeted
Run Flint locally on 24GB RAM via Unsloth Dynamic GGUFs!
Flint is the strongest MoE model under 50B parameters.
Recursive’s Apache-2.0 model has 39.8B parameters (2.5B active). Run at 125 t/s on Mac, 150 on RTX Spark
GGUF: huggingface.co/unsloth/flint…
Guide: unsloth.ai/docs/models/flint
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.
Tim Rocktäschel retweeted
Today, we release our first model, Flint-40B-A2B, a fast local coding model built with our automated research system. huggingface.co/RecursiveSI/f…
1. Strong performance. 54.9% on Terminal-Bench 2.1, 89.7% on SWE-bench Pro v2, and 74.2% on SWE-bench Verified.
2. It is trained from scratch (tokenizer->Pretrain->Midtrain->SFT->RL).
3. Fast in your local laptop. Up to 2x faster than comparable models: 125+ tokens/sec on a MacBook Pro, 150+ on RTX Spark.
4. We build the complete model training pipeline in 4 month!
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.
Tim Rocktäschel retweeted
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.
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.
Tim Rocktäschel retweeted
88.8 tok/s on an M5 Pro over an 11.9k-token generation, via LM Studio, fully offline. And the quality is really good, better than I expected from a model this fast. Congrats to the @Recursive_SI team!
Tim Rocktäschel retweeted
Evolution created the 20W human brain through open-ended search, pushing compute to the edge so latency wouldn't let entropy win.
@Recursive_SI is running accelerated evolution. Instead of hiring more researchers, their Eureka Machine built Flint in under 4 months.
wired.com/story/recursive-se…
Tim Rocktäschel retweeted
Tested on MacBook Pro M5 MAX 64GB
- Model: Recursive Flint (40B MoE, 2.5B active), int4 MLX via rec, in pi
- TPS: 114–132 output tok/s
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.
Tim Rocktäschel retweeted
Excited to support for @Recursive_SI Flint 🚀
40B parameters, just 2.5B active per token, built for fast local coding inference.
We’ve tested it with OpenHands. Now developers have another open model for real engineering workflows.
openhands.dev/blog/recursive…
Tim Rocktäschel retweeted
Flint from @Recursive_SI is now on OpenRouter!
Recursive's first model: an open-weight 40B MoE (~2.5B active) built for coding agents by their autoresearch system, the Eureka Machine. 262K context, reasoning and tool calling.
Recursive reports 54.9% on Terminal-Bench 2.1, 89.7% on SWE-bench Pro v2 and 74.2% on SWE-bench Verified.
Free to use at launch.
Try it now: openrouter.ai/recursive/flin…
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.
Get started in two lines of code.
Fast is fast! Here on an M5 Max with pi + rec, their inference engine. Like this trend of model providers offering dedicated engine too!
As easy as:
curl -fsSL raw.githubusercontent.com/re… | bash
rec launch pi
Repo here: github.com/recursive-oss/rec
Tim Rocktäschel retweeted
Recursive's Flint at a glance 👇 40B-A2.5B coding MoE trained from scratch:
– 256 experts (512 wide), top-6 + 2 shared (512 x 2), sigmoid router
– 3:1 sliding-window (513): NoPE global attention
– learned attention sinks, QK-norm, 32q/4kv GQA
– NVFP4 incl. a 4-bit lm_head (12% of active weights), self-distilled to within 0.7 pts of BF16
– ~7 GiB KV at 262K ctx
– 131 tok/s on M4 Pro (69 avg to 128K)
Weights: huggingface.co/RecursiveSI/f…
Thinking of making these one-pagers a tradition for new model releases 🤗
Tim Rocktäschel retweeted
@Recursive_SI shipped its first model, Flint! Exciting to see they did extensive RL training on the Pi harness, so it works well out of the box.
40B MoE (2.5B active) for fast local coding. Apache 2.0.
Available now for free in Pi via Radius.
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.
Tim Rocktäschel retweeted
We’re about to see a breakout in local model capabilities.
Flint scores 54.9% on Terminal-Bench 2.1 with ~2B active parameters. That’s ahead of every other small MoE and two of the three dense ~30B models.
It only needs about 24GB and runs at 125+ tok/s on a MacBook Pro.
Cool application of @Recursive_SI's Flint for autoresearch—locally at home.
Tim Rocktäschel retweeted
We’ve partnered with @Recursive_SI to bring day-0 support for Flint, their first OSS model, to Apple Silicon with @Nativ_AI and mlx-vlm 🚀
Key features:
→ 40B params, 2.5B active
→ Apache 2.0
→ Runs 100% local on Apple Silicon
🧵
Tim Rocktäschel retweeted
My bar for AGI is still an AI winning a Nobel.
But an AI researcher looking at the lm_head (that *everyone* tells you to keep in high precision) deciding to quantize it to 4 bits AND being actually right (0.7pt from bf16) is definitely impressive.
Even better when the result are open-weights pushed on the Hugging Gace hub.
Congrats @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 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.
Tim Rocktäschel retweeted
SLIDING WINDOW AGAIN
and huge congrats, this is a very cool release :)
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.