Flint, Recursive’s first public model, showcases a bold leap toward self‑improving AI by running entirely on a laptop. Built largely by Recursive’s own automated research system, the Eureka Machine, Flint delivers coding assistance without cloud dependence, offering developers privacy and zero usage fees.
Building Flint with the Eureka Machine
The Eureka Machine guided nearly every phase of Flint’s creation—from pre‑training to inference—allowing a small team to deliver a 40‑billion‑parameter model in under four months. By employing low‑bit training and a sparser model shape, Recursive shrank Flint’s size by 71% while preserving quality, and custom kernels accelerated its computations. The result is a mixture‑of‑experts architecture that activates only 2.5 billion parameters per task, making the model lightweight enough for ordinary laptops.
- 40 billion total parameters, 2.5 billion active per task
- 71% size reduction via low‑bit training
- Runs locally on laptops, no cloud required
- Apache 2.0 licensed, free to download on Hugging Face
- 131 tokens/second on MacBook Pro M4 Pro; 150 tokens/second on RTX Spark
Why Flint Matters for Local AI
Flint’s local deployment eliminates the need for costly cloud inference, enabling developers to keep code private and avoid per‑token fees. Its open‑source license invites community contributions and custom modifications, fostering broader experimentation. Early reviews, such as Wired’s Will Knight, praised Flint’s reliability on small projects, noting it outperformed larger paid assistants on specific tasks.
Performance & Benchmarks
| Model | Parameters | Tokens/s (M4 Pro) | SWE‑Bench Score |
|---|---|---|---|
| Flint | 40 B (2.5 B active) | 131 | Highest among <50 B models |
| Google Gemma 4 | 10 B | 58‑69 | Lower |
| Nvidia Nemotron 3.5 Lightning | 10 B | ~60 | Lower |
| Meta Muse Spark Open | 816 B | N/A (requires >1 TB RAM) | Not comparable |
| Alibaba Qwen3.8 27B | 27 B | N/A | Only model scoring higher than Flint on our internal tests |
Charting the Path to Recursive Self‑Improvement
How does Flint’s mixture‑of‑experts architecture enable it to run on a laptop?
Flint’s mixture‑of‑experts design activates a small, task‑specific subset of its full parameter set—just 2.5 billion out of 40 billion—during inference. This selective activation drastically reduces memory and compute requirements, allowing the model to fit within the RAM constraints of a typical laptop. By limiting the active parameters, the system also speeds up token generation, achieving 131 tokens per second on an M4 Pro chip.
What role did Recursive’s Eureka Machine play in Flint’s development and performance gains?
The Eureka Machine orchestrated nearly every development step, from generating training data to fine‑tuning hyperparameters. Its automated experimentation loop identified optimal sparse architectures and low‑bit training regimes, cutting the model’s size by 71% without sacrificing quality. Additionally, the system crafted custom kernels that accelerated core operations, contributing to Flint’s superior token throughput compared to peer models.
How does Flint’s token generation speed compare to other leading coding models on the same hardware?
On a MacBook Pro with Apple’s M4 Pro chip, Flint achieved 131 tokens per second—roughly double the 58‑69 tokens per second produced by Google Gemma 4 and Nvidia Nemotron 3.5 Lightning. On Nvidia’s RTX Spark, Flint topped 150 tokens per second, surpassing its competitors by a significant margin. These speed advantages stem from Flint’s efficient mixture‑of‑experts architecture and custom kernel optimizations.