Instructions to use SandLogicTechnologies/clif-GGUF with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- llama.cpp
How to use SandLogicTechnologies/clif-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/clif-GGUF:IQ3_M # Run inference directly in the terminal: llama cli -hf SandLogicTechnologies/clif-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/clif-GGUF:IQ3_M # Run inference directly in the terminal: llama cli -hf SandLogicTechnologies/clif-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/clif-GGUF:IQ3_M # Run inference directly in the terminal: ./build/bin/llama cli -hf SandLogicTechnologies/clif-GGUF:IQ3_M
Use Docker
docker model run hf.co/SandLogicTechnologies/clif-GGUF:IQ3_M
- Llama new
- LM Studio
- Jan
- Ollama
How to use SandLogicTechnologies/clif-GGUF with Ollama:
ollama run hf.co/SandLogicTechnologies/clif-GGUF:IQ3_M
- Unsloth Desktop
- Docker Model Runner
How to use SandLogicTechnologies/clif-GGUF with Docker Model Runner:
docker model run hf.co/SandLogicTechnologies/clif-GGUF:IQ3_M
- Lemonade
How to use SandLogicTechnologies/clif-GGUF with Lemonade:
Pull the model
# Download Lemonade from https://lemonade-server.ai/ lemonade pull SandLogicTechnologies/clif-GGUF:IQ3_M
Run and chat with the model
lemonade run user.clif-GGUF-IQ3_M
List all available models
lemonade list
- Atomic Chat
Clef
Clef is a 27B multimodal decision model developed by Cloudflare. It turns a state and a schema of typed questions into decisions. The state can be text, JSON, images or video, and the model returns a probability for every allowed option of every question in a single forward pass.
Unlike a chat model, Clef does not generate free-form text or require output parsing. Each question is answered as a true/false decision, a choice among named options, or a score over ordered options, which makes its output directly usable in business workflows such as routing, triage and invoice processing.
This repository provides GGUF quantized versions of Clef for local inference, built from an F16 intermediate conversion to reduce storage and memory requirements.
Model Overview
- Model Name: Clef
- Base Model: Cloudflare/clef (post-trained from Qwen/Qwen3.8-27B)
- Architecture: Multimodal Transformer backbone with a joint schema head
- Parameter Count: 27 Billion Parameters
- Modalities: Text, JSON, Image, Video
- Primary Language: English
- Developer: Cloudflare
- License: Apache 2.0
Quantization Formats
This repository provides multiple GGUF quantized versions of Clef designed to reduce storage and memory requirements compared to the 16-bit conversion. The available formats offer different trade-offs between size and numerical precision.
IQ3_M
- Size reduction of approx. 76.74% (11.71 GB) compared to 16-bit (50.35 GB)
- Aggressive 3-bit quantization providing the smallest GGUF representation among the supplied conversions
- Suitable for constrained local environments where model footprint is the primary consideration
- The lower numerical precision introduces a greater trade-off for close decisions and multi-criteria questions
IQ4_NL
- Size reduction of approx. 70.59% (14.81 GB) compared to 16-bit (50.35 GB)
- 4-bit non-linear quantization that retains more weight information than the 3-bit variant
- Suitable for general local deployment, offering a practical balance between size and decision quality
- Provides a good default choice for routing, classification and typed-decision workloads
IQ4_XS
- Size reduction of approx. 71.92% (14.14 GB) compared to 16-bit (50.35 GB)
- Compact 4-bit representation, slightly smaller than IQ4_NL
- Suitable for deployments with tighter storage or memory constraints while keeping a 4-bit weight format
Q6_K
- Size reduction of approx. 58.97% (20.66 GB) compared to 16-bit (50.35 GB)
- Higher-precision 6-bit K-Quant representation that retains more numerical information than the 4-bit and 3-bit variants
- Provides the largest GGUF file among the supplied quantized versions
- Suitable for decisions where accuracy on close or borderline cases matters most
Training Background (Original Model)
Clef is post-trained from Qwen/Qwen3.8-27B and is built around typed decision-making rather than general conversation. Its design focuses on turning structured and multimodal state into reliable decisions over a defined schema.
Typed Decision Architecture
- Built on a Qwen3.8-27B backbone, including its vision encoder
- Uses a joint schema head that reads the backbone's final hidden states
- Scores all options of all questions in one pass, rather than generating an answer as text
- Supports three question types: true/false (
noul), named choice (choice), and ordered score (score)
Key Capabilities
Typed Decisions
Answers true/false, choice, and score questions over a given state, returning probabilities for each option.Multimodal State Understanding
Reads state as text, JSON, images, or video, so a single schema can cover documents, messages, and visual evidence.Structured Output Without Parsing
Returns logits per option, so no free-form generation or output parsing is required.Tool and Routing Selection
Chooses the relevant tool, team, or category for a request, with strong results on BFCL.Business Workflow Decisions
Supports decisions such as invoice actions, customer service handling, and security incident triage.Batched Mixed Inputs
Text-only and multimodal records can be processed in the same batch.Long-Context Input
Accepts long states, with a default limit of 16,384 tokens per encoded record that can be adjusted.
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\clef-IQ4_NL.gguf \
-p "State: Invoice 1042 from Acme for 1250 USD is 45 days past due. Question: Is this invoice overdue? Answer true or false."
Recommended Usecases
Ticket and Message Routing
Route support messages to the right team based on the content and urgency.Invoice and Document Processing
Decide actions such as paid, overdue, or needs review from structured records.Security Incident Triage
Classify incidents and decide on the appropriate response from logs and reports.Tool Selection for Agents
Choose the correct tool or function for a user request inside an agent workflow.Multimodal Review
Answer true/false or choice questions about attached images or video frames, such as whether a receipt total is legible.Classification and Moderation Pipelines
Assign labels or scores to content where a fixed set of options is required.
Clef is intended primarily for English-language decision workflows over a defined schema. It returns probabilities, so downstream systems should apply their own thresholds and human review for high-impact decisions.
Acknowledgments
These quantized models are based on the original work by the Cloudflare team.
Special thanks to:
The Cloudflare team for developing and releasing Clef and the joint schema decision model.
Georgi Gerganov and the
llama.cppopen-source community for enabling efficient local inference and GGUF-based model deployment.
Contact
For questions, feedback, or support, please reach out at support@sandlogic.com or visit https://www.sandlogic.com/.
- Downloads last month
- 594
3-bit
4-bit
6-bit