Open LLM Deploy

Recursive Flint brings a quantized 40B coding MoE into local deployment discussions

Recursive Flint brings a quantized 40B coding MoE into local deployment discussions

Recursive Flint is reported as an Apache-2.0 40B-total/2.5B-active coding MoE, with an approximately 80 GB BF16 checkpoint and local use dependent on quantization. Claims include 125–131 tokens/s on M4 Pro, 74.2% SWE-bench, and 262K context; independent reproduction and long-context memory/quality tests remain needed.

Sources (1)
Updated Nov 8, 2026