One researcher pretrains 100B-500B coding model in 3.5 months on under $10M, agents write 99% of code
RobertTLange · x · 2026-10-22
RobertTLange highlights two landmark recursive self-improvement releases. Mirendil-Base: one researcher pretrained a 100B–500B coding model from scratch in 3.5 months with under $10M of compute, with agents writing over 99% of the code and running experiments while the human set direction. Larger versions are already training. Recursive also shipped Flint, an open-weights 40B local coding model built in under 4 months by its automated research system (Eureka Machine), running 125+ tok/s on an M4 Pro. The two differ in framing: Mirendil compresses people/compute/time per model, Recursive proves automated research can ship a runnable model.
More from coding & agent
- Dev builds a cookie clicker for humans and AI on ChatGPT Sites with database and MCP — craigsdennis · 2026-10-22
- Dev builds a human-and-AI Cookie Clicker with ChatGPT Sites, a database and an MCP server — craigsdennis · 2026-10-22
- Start agent factory automation: one idle issue per day, iterate on quality — intellectronica · 2026-10-22
- Recursive releases Flint: 40B MoE local coding model, built in 4 months — _rockt · 2026-10-22
- Hamel Husain and Shreya Shankar Launch 'Evals for AI Engineers' Book Nov 12 — HamelHusain · 2026-10-22
- EAU meta-agent automates RL training: 36.7% higher task completion on ColBench — xiye_nlp · 2026-10-22