Recursive releases Flint, 40B model for local coding
Flint activates just 2.5 billion of its 40 billion parameters per token, reducing the compute needed to run coding agents locally.
Recursive reports speeds above 125 tokens per second on an M4 Pro MacBook Pro and 150 on a 64GB NVIDIA RTX Spark at short contexts using its optimized engine. The model supports 512k tokens of context, though the company recommends staying within 256k. Recursive advises breaking complex jobs into smaller subtasks and giving explicit instructions, acknowledging that Flint may need more guidance than larger closed-source models. Its Apache 2.0 weights work with tools including llama.cpp, Ollama and LM Studio, allowing offline use with data kept on-device.
TestingCatalog's note
Recursive reports speeds above 125 tokens per second on an M4 Pro MacBook Pro and 150 on a 64GB NVIDIA RTX Spark at short contexts using its optimized engine. The model supports 512k tokens of context, though the company recommends staying within 256k. Recursive advises breaking complex jobs into smaller subtasks and giving explicit instructions, acknowledging that Flint may need more guidance than larger closed-source models. Its Apache 2.0 weights work with tools including llama.cpp, Ollama and LM Studio, allowing offline use with data kept on-device.
TestingCatalog's note
score 5