Instructions to use SandLogicTechnologies/flint-GGUF with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use SandLogicTechnologies/flint-GGUF with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf SandLogicTechnologies/flint-GGUF:IQ3_M
# Run inference directly in the terminal:
llama cli -hf SandLogicTechnologies/flint-GGUF:IQ3_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf SandLogicTechnologies/flint-GGUF:IQ3_M
# Run inference directly in the terminal:
llama cli -hf SandLogicTechnologies/flint-GGUF:IQ3_M

Build from source code

git clone https://github.com/ggml-org/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-app
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama serve -hf SandLogicTechnologies/flint-GGUF:IQ3_M
# Run inference directly in the terminal:
./build/bin/llama cli -hf SandLogicTechnologies/flint-GGUF:IQ3_M

Use Docker

docker model run hf.co/SandLogicTechnologies/flint-GGUF:IQ3_M

Llama new
LM Studio
Jan

vLLM

How to use SandLogicTechnologies/flint-GGUF with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "SandLogicTechnologies/flint-GGUF"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "SandLogicTechnologies/flint-GGUF",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/SandLogicTechnologies/flint-GGUF:IQ3_M

Ollama
How to use SandLogicTechnologies/flint-GGUF with Ollama:
```
ollama run hf.co/SandLogicTechnologies/flint-GGUF:IQ3_M
```
Unsloth Desktop

How to use SandLogicTechnologies/flint-GGUF with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf SandLogicTechnologies/flint-GGUF:IQ3_M

Configure the model in Pi

# Install Pi:
npm install -g @earendil-works/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "SandLogicTechnologies/flint-GGUF:IQ3_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use SandLogicTechnologies/flint-GGUF with Docker Model Runner:
```
docker model run hf.co/SandLogicTechnologies/flint-GGUF:IQ3_M
```

Lemonade

How to use SandLogicTechnologies/flint-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull SandLogicTechnologies/flint-GGUF:IQ3_M

Run and chat with the model

lemonade run user.flint-GGUF-IQ3_M

List all available models

lemonade list

Hermes Agent

How to use SandLogicTechnologies/flint-GGUF with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf SandLogicTechnologies/flint-GGUF:IQ3_M

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default SandLogicTechnologies/flint-GGUF:IQ3_M

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use SandLogicTechnologies/flint-GGUF with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf SandLogicTechnologies/flint-GGUF:IQ3_M

Configure OpenClaw

# Install OpenClaw:
npm install -g openclaw@latest
# Register the local server and set it as the default model:
openclaw onboard --non-interactive --mode local \
  --auth-choice custom-api-key \
  --custom-base-url http://127.0.0.1:8080/v1 \
  --custom-model-id "SandLogicTechnologies/flint-GGUF:IQ3_M" \
  --custom-provider-id llama-cpp \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

Recursive Flint

Recursive Flint is a mixture-of-experts language model developed by Recursive for repository-level software engineering and coding-agent workloads. It is designed to investigate codebases, diagnose and fix software issues, implement new functionality, and create or modify tests through iterative reasoning and tool interaction.

Flint contains approximately 39.8 billion total parameters with around 2.5 billion active parameters per token. Its architecture combines sparse MoE layers with hybrid attention, using sliding-window attention in most layers and full attention in selected layers to support both local code understanding and long-range repository reasoning.

The GGUF variants in this repository provide compressed representations of Flint for local inference. The available quantizations significantly reduce the storage footprint of the original BF16 conversion while retaining the model's coding, reasoning, and tool-use capabilities.

Model Overview

Model Name: Recursive Flint
Base Model: RecursiveSI/flint
Parameter Count: Approximately 39.8 Billion Parameters
Active Parameters: Approximately 2.5 Billion Parameters per Token
Architecture: Decoder-only Mixture-of-Experts Transformer
Layers: 49
Experts: 256 routed experts with 6 selected per token, plus a shared expert
Attention: Hybrid sliding-window and full attention
Context Length: 262,144 tokens
Attention Heads: 32 query heads and 4 key/value heads
Modalities: Text
Primary Focus: Repository-Level Software Engineering
Developer: Recursive
License: Apache 2.0

Flint uses two dense layers followed by sparse MoE layers. Of its 49 layers, 35 use sliding-window attention with a 513-token window, while 14 use full attention. This hybrid design combines local code-context processing with selected full-context layers for broader repository reasoning.

Quantization Formats

This repository provides multiple GGUF quantized versions of Recursive Flint designed to reduce the storage and memory requirements of the BF16 model. The supplied formats offer different trade-offs between compression and numerical precision.

IQ4_NL

Size reduction of approx. 71.49% (21.14 GB) compared to 16-bit (74.16 GB)
4-bit non-linear quantization providing a substantially smaller representation of the Flint model
Retains more weight information than the supplied 3-bit variant while significantly reducing the model footprint
Suitable for coding-agent workloads involving repository analysis, debugging, code generation, and tool-assisted development
Provides a practical balance between local deployment requirements and numerical representation for demanding software-engineering tasks

IQ4_XS

Size reduction of approx. 73.00% (20.02 GB) compared to 16-bit (74.16 GB)
Compact 4-bit representation intended for deployments with tighter storage and memory constraints
Produces a smaller GGUF file than IQ4_NL while remaining within the same general 4-bit quantization class
Suitable for repository-level coding, codebase investigation, debugging, and local coding-agent experimentation
Provides an alternative when reducing model footprint is preferred while maintaining a 4-bit representation

IQ3_M

Size reduction of approx. 78.01% (16.31 GB) compared to 16-bit (74.16 GB)
Aggressive 3-bit quantization providing the smallest GGUF representation among the supplied conversions
Significantly reduces storage and memory requirements for the 39.8B-parameter MoE model
Suitable for constrained local environments where model footprint is a primary deployment consideration
The lower numerical precision introduces a stronger trade-off for difficult reasoning, complex debugging, and long-horizon software-engineering tasks

Training Background (Original Model)

Flint is developed specifically around agentic software engineering rather than general-purpose conversational assistance. Its training and evaluation focus on tasks that require understanding an existing repository, reasoning over implementation details, making code changes, executing tools, and validating the resulting changes.

Repository-Level Reasoning

Designed to investigate and understand existing software repositories
Supports analysis of issues that can span multiple files and implementation components
Performs reasoning over code structure, dependencies, and relevant repository context
Targets software-engineering tasks such as bug fixing and feature implementation
Supports long-horizon workflows where multiple reasoning and tool-use steps are required

Hybrid MoE Architecture

Uses 39.8B total parameters with approximately 2.5B active parameters per token
Contains 49 Transformer layers
Uses 256 routed experts in sparse MoE layers
Selects 6 routed experts per token together with an always-active shared expert
Combines sliding-window attention with full-attention layers for different context ranges

Agentic Tool Use

Supports multi-turn tool interaction as part of coding workflows
Can be integrated with coding-agent environments that provide repository and execution tools
Designed to investigate code, modify implementations, and validate changes through executable workflows
Supports reasoning during tool-driven software-engineering tasks
Evaluated on agentic coding benchmarks including SWE-bench Pro, SWE-bench Verified, and Terminal-Bench 2.1.

Key Capabilities

Repository-Level Software Engineering
Analyzes existing repositories and works through software-engineering tasks that require understanding the surrounding codebase.
Codebase Investigation
Helps identify relevant files, implementation paths, dependencies, and code components associated with a task or reported issue.
Bug Fixing
Supports diagnosing software problems and developing candidate changes intended to resolve repository-level issues.
Feature Implementation
Can reason about requested functionality and generate code changes spanning the components required to implement a feature.
Multi-File Code Reasoning
Handles tasks where understanding and modifying several related files is necessary to complete the requested change.
Test Generation and Execution
Supports software-engineering workflows that involve writing tests and using test results to validate code modifications.
Tool-Assisted Coding
Supports multi-turn interaction with external tools as part of an iterative coding-agent workflow.
Long-Context Code Reasoning
Provides a documented context configuration of up to 262,144 tokens, enabling analysis of large amounts of repository and task information.

Usage Example

Using llama.cpp

The GGUF model can be loaded through a compatible llama.cpp build for local text inference.

./llama-cli \
  -m SandlogicTechnologies/flint-IQ4_NL.gguf \
  -p "Inspect this repository issue, identify the likely source of the problem, propose the required code changes, and explain how the fix should be tested."

Recommended Usecases

Repository-Level Coding Agents
Build automated software-engineering systems capable of investigating and modifying existing codebases.
Bug Diagnosis and Repair
Analyze repository issues, identify likely causes, and generate candidate fixes.
Feature Development
Assist with implementing new functionality across existing software projects.
Code Maintenance and Refactoring
Support changes that require coordinated modifications across multiple files and components.
Test-Assisted Development
Integrate code generation with test creation, execution, and iterative debugging.
Tool-Using Coding Agents
Build agents that combine reasoning with repository tools and execution environments.
Software Engineering Research
Experiment with long-horizon coding tasks, repository reasoning, and agentic software development.
Local Coding Assistance
Deploy compressed GGUF variants for local experimentation and development when the BF16 representation is too large for the available environment.

Flint is primarily intended for text-based software engineering and coding-agent workflows. Its official evaluation focuses on repository-level engineering and agentic coding rather than multimodal tasks.

Acknowledgments

These quantized models are based on the original work by the Recursive development team.

Special thanks to:

The Recursive team for developing and releasing Flint and for advancing research into agentic software engineering and repository-level coding.
The open-source software-engineering community for developing coding-agent frameworks, evaluation environments, and tooling used to investigate repository-level model capabilities.
Georgi Gerganov and the llama.cpp open-source community for enabling efficient GGUF conversion and local model inference.

Contact

For questions, feedback, or support, please reach out at support@sandlogic.com or visit https://www.sandlogic.com/.

Downloads last month: 79

GGUF

Model size

40B params

Architecture

flint

Hardware compatibility

3-bit

4-bit

Model tree for SandLogicTechnologies/flint-GGUF

Base model

RecursiveSI/flint

Quantized

(96)

this model