Recursive's Flint: 40B MoE local coding model, 125+ tok/s on M4 Pro, built in 4 months by automation
josh_tobin_ · x · 2026-10-22
Recursive shipped its first model Flint less than 6 months after starting the company: a 40B-parameter MoE (2.5B active per token) agentic coding model meant to run locally or self-hosted, offloading everyday coding from frontier models.
Key facts:
- 125+ tokens/s on M4 Pro MacBook Pro, 150 tokens/s on RTX Spark; leads MoE models under 50B total parameters on agentic coding benchmarks
- Apache 2.0 open weights on Hugging Face, runnable in OpenCode and Pidotdev
- vLLM collaboration cut kernel launches from 22 to 7 per layer, enabling 500 tokens/s on a single GB200 with 100K context
- Built from scratch in under 4 months by the Eureka Machine, an automated research system that designed training recipes and inference kernels
CEO Josh Tobin argues everyday coding will shift to models running on your laptop or a nearby server, for privacy and private, unmetered intelligence.
More from coding & agent
- Migrating an entire Hermes automation workload to Antigravity took about 2 hours — rseroter · 2026-10-22
- How a second Mac mini lets AI agents auto-distribute compute-heavy tasks — tedddyoweh · 2026-10-22
- LangChain's Design Agent Shipped 3,000+ Assets and Earned 4.5M Views Since August — LangChain · 2026-10-22
- Databricks Chief AI Scientist: Evals Are the Biggest Bottleneck to Enterprise AI Adoption — jefrankle · 2026-10-22
- Developer plans a solar-aware job scheduler so his homelab rides free excess power — mertdumenci · 2026-10-22
- How do you design your agentic code review workflow? — lucasmeijer · 2026-10-22