Recursive launches Flint: open-weight local coding model hitting 125+ tokens/sec on MacBook Pro
_rockt · x · 2026-10-23
Recursive AI (Richard Socher) introduced Flint, its first model: a fast local coding model built from scratch in under 4 months by its automated research system, the Eureka Machine.
Key facts:
- 125+ tokens/sec on a MacBook Pro and 150+ on RTX Spark — claimed up to 2x faster than comparable models
- Open weights on Hugging Face; runnable locally or free via opencode and pidotdev
- Positioning: private, unmetered intelligence on your own machine
The post also includes Atomic Agent's local agent benchmark: on GAIA Level 1 (53 real-world tasks, same model and machine), Atomic Agent scored 69.8% vs Hermes Agent's 58.5%, at 217s vs 351s per task (1.6x faster). Techniques include GBNF grammar-constrained decoding for structurally valid tool calls, KV-cache prefix reuse, and parallel read-only batches.
More from coding & agent
- Dev adds local MCP server to Mac screenshot app so Claude designs from your references — Embarrassed-Ebb-740 · 2026-10-23
- Open Duet, a self-hosted open-source Claude Duet alternative, hits 6.1k GitHub stars — matchaman11 · 2026-10-23
- Hermes Agent turns into an OSINT situation room with live 3D globe and auto sitreps — Teknium · 2026-10-23
- IQuest-Q1 debuts: 320B sparse MoE foundation model for agentic CLI coding — iQuestResearch · 2026-10-23
- 'This is literally Jarvis': voice-driven ChatGPT designs CAD models in FreeCAD and 3D prints them — granawkins · 2026-10-23
- New US federal rule requires contractors to report AI models in 3 days; endpoint abstraction is the cheapest fix — AIFlow_ML · 2026-10-23