Description
Flint is a sophisticated mixture-of-experts language model developed by Recursive, specifically tailored for coding agents. With a total of 39.8 billion parameters and approximately 2.5 billion active per token, Flint is engineered to perform a variety of tasks including investigating codebases, fixing bugs, implementing features, and writing tests. The model supports text input and output, reasoning, and multi-turn tool use, making it a versatile tool for developers.
Read More
In terms of performance, Flint scores 54.9% on Terminal-Bench 2.1, 89.7% on SWE-bench Pro v2, and 74.2% on SWE-bench Verified. These scores reflect its capability in handling complex coding tasks. The model is built on a decoder-only mixture-of-experts Transformer architecture, featuring 49 layers, 35 sliding-window layers, and 14 full-attention layers. It employs a sliding window of 513 tokens and utilizes 32 query heads and 4 key/value heads for attention.
Flint is available as an open-source model under the Apache 2.0 license, with a BF16 checkpoint, model configuration, tokenizer, chat template, and an SGLang serving kit included in the repository. Quantized variants are available for NVIDIA and Apple Silicon runtimes. The model's training involved pretraining, mid-training, supervised fine-tuning, and reinforcement learning, covering text and code, instruction following, and agentic tasks.
Despite its capabilities, Flint has limitations. It can produce incorrect or harmful content, faulty code, or inappropriate tool calls. Users are advised to test generated code and review changes before implementation. The model's performance can vary based on checkpoint precision and inference runtime, and long reasoning tasks may increase latency. For support, users can contact Recursive with specific details about their setup.
- CategoryAI Models & LLMs
- Pricing ModelOpen source
- DeploymentCloud
- PlatformsWeb
Flint Highlights
Mixture-of-experts language model
39.8 billion total parameters
2.5 billion active parameters per token
Supports text input and output
Reasoning and multi-turn tool use
Open-source under Apache 2.0 license
Quantized variants for NVIDIA and Apple Silicon
Supervised fine-tuning and reinforcement learning
Getting Started with Flint
Access model page: Visit the Flint repository on Hugging Face
Authenticate: Ensure you have the necessary credentials
Configure: Set up the model using the provided configuration files
Send prompts: Use the chat template to interact with the model
Fine-tune: Adjust the model settings for specific tasks
Flint's Use Cases
- Codebase Investigation
- Bug Fixing
- Feature Implementation
- Test Writing
- Multi-Turn Tool Use
