Recursive has introduced Flint, a decoder-only MoE Transformer model designed for coding agents. The model supports reasoning and OpenAI-compatible tool calls, with capabilities including codebase analysis, bug fixing, feature implementation, and test writing.
The architecture consists of 49 layers, including 2 dense layers and 47 MoE layers. Each MoE layer features 256 routed experts, with 6 experts selected per token alongside one shared expert. The model utilizes a mix of 35 sliding window attention layers and 14 full attention layers.
Flint operates with 2.5 billion active parameters, including the output head, while the active body contains approximately 2.2 billion parameters. The model was developed using the Eureka Machine, an automated research system that handles idea proposal, implementation, and testing.
Performance metrics indicate speeds exceeding 125 tokens per second on an M4 Pro MacBook Pro and over 150 tokens per second on a 64 GB NVIDIA RTX Spark. Recursive claims these speeds are up to 2x faster than comparable models and over 5x faster than dense 25-40B models with similar coding scores.
