Recursive Releases Flint, a 40B MoE Coding Model
Recursive introduces Flint, an open-weight 40B MoE local coding model with 2.5B active parameters, hitting 74.2% on SWE-bench Verified and 125+ tokens per second.
Stock photo for illustration only, not from the actual event
- Recursive has launched Flint, an open-weight local coding model licensed under Apache 2.0.
- It features a 40B total parameter count while utilizing only 2.5B active parameters per token.
- The model achieves 74.2% on SWE-bench Verified and over 125 tokens per second on an M4 Pro chip.
Artificial intelligence developer Recursive has officially released Flint, a local coding model designed specifically to empower AI agents directly on consumer hardware. Distributed under the permissive Apache 2.0 license, this open-weight release gives developers full freedom to integrate and deploy the technology locally without cloud dependency.
The core architectural highlight of Flint lies in its Mixture of Experts (MoE) design, boasting a total of 40 billion parameters while activating only 2.5 billion active parameters during inference. This clever engineering approach drastically reduces hardware memory bandwidth and compute requirements while preserving the deep contextual knowledge of a much larger model.
Stock photo for illustration only, not from the actual event
When put through rigorous evaluation benchmarks, Flint secured an impressive 74.2% score on SWE-bench Verified. Furthermore, real-world benchmark tests running on an Apple MacBook Pro powered by the M4 Pro processor demonstrated an exceptional text-generation speed exceeding 125 tokens per second.
The industry shift toward local MoE coding models like Flint highlights a growing demand for privacy-focused, low-latency developer tools. By keeping active parameters down to 2.5B within a 40B framework, developers can leverage high-performance AI agent capabilities locally on powerful consumer silicon like Apple's M-series chips, bypassing the latency and data privacy concerns associated with remote cloud APIs.
Source: MarkTechPost
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment