Recursive Launches Flint, Free Laptop Coding AI Built by Its Eureka Machine
Recursive's free Flint coding model runs on laptops, beats sub-50B rivals, and was built in under four months by the Eureka Machine AI system in 2026.
Summary
Recursive, founded in 2025 and led by CEO Richard Socher, Salesforce’s former chief scientist, released its first model, Flint, on October 21, 2026. The free Apache 2.0 coding model runs locally and offline on ordinary laptops, protecting code and avoiding token-based cloud fees, and is downloadable from Hugging Face. Recursive says its human-supervised Eureka Machine proposed ideas, ran experiments and selected next steps across pre-training, post-training and inference, helping a small team build Flint from scratch in under four months. Its contributions included a sparser architecture, custom kernels and low-bit training that cut model size 71% with almost no quality loss. Recursive emerged from stealth in May with $650 million from GV, Nvidia and AMD Ventures at a $4.65 billion valuation, then signed a $410 million Amazon Web Services compute deal in June focused on agent count rather than headcount.
Flint is a 40 billion parameter mixture-of-experts model activating 2.5 billion parameters per task. It generated about 131 tokens per second on an Apple M4 Pro MacBook Pro, versus 58 to 69 for comparable models including Google’s Gemma 4 and Nvidia’s Nemotron 3.5 Lightning, and exceeded 150 on Nvidia’s RTX Spark. Meta’s week-old Muse Spark Open uses 816 billion parameters and requires at least 1TB of memory. Flint led all models below 50 billion parameters on SWE-Bench Verified; across Recursive’s six-test average, only Alibaba’s Qwen3.8 27B scored higher. Will Knight used a 48GB Mac to build a to-do app and web research tool, with every request succeeding despite common small-model errors or lag versus Claude and Codex. Flint still needs more guidance than large paid assistants, especially task decomposition. Recursive’s target is stoppable, human-overseen recursive self-improvement, where each model improves Eureka and extends autonomous work. OpenAI said the previous month that by mid-August its agents logged 3.1 workdays per human researcher day.
Positives
- Flint generated 131 tokens per second on an M4 Pro MacBook Pro and exceeded 150 on Nvidia’s RTX Spark.
- Low-bit training reduced Flint’s size by 71% with almost no quality loss.
- Flint led every model below 50 billion parameters on SWE-Bench Verified.
- Apache 2.0 licensing lets developers download, use and modify Flint free through Hugging Face.
- Local, offline operation protects developers’ code and avoids usage-based cloud fees.
- The Eureka Machine helped a small team build Flint from scratch in under four months.
Risks & concerns
- Flint requires more guidance than large paid coding assistants, including breaking substantial jobs into smaller tasks.
- Alibaba’s Qwen3.8 27B beat Flint across Recursive’s six-test average.
- Recursive did not disclose how much researcher time the Eureka Machine saved.
- Recursive self-improvement remains controversial because increasingly autonomous AI could become harder for humans to control.