Instructions to use Agentware/d3-flash-calibrated-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 Agentware/d3-flash-calibrated-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 Agentware/d3-flash-calibrated-gguf:F32
# Run inference directly in the terminal:
llama cli -hf Agentware/d3-flash-calibrated-gguf:F32

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf Agentware/d3-flash-calibrated-gguf:F32
# Run inference directly in the terminal:
llama cli -hf Agentware/d3-flash-calibrated-gguf:F32

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 Agentware/d3-flash-calibrated-gguf:F32
# Run inference directly in the terminal:
./build/bin/llama cli -hf Agentware/d3-flash-calibrated-gguf:F32

Use Docker

docker model run hf.co/Agentware/d3-flash-calibrated-gguf:F32

Llama new
LM Studio
Jan
Ollama
How to use Agentware/d3-flash-calibrated-gguf with Ollama:
```
ollama run hf.co/Agentware/d3-flash-calibrated-gguf:F32
```
Unsloth Desktop

How to use Agentware/d3-flash-calibrated-gguf with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Agentware/d3-flash-calibrated-gguf:F32

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": "Agentware/d3-flash-calibrated-gguf:F32"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use Agentware/d3-flash-calibrated-gguf with Docker Model Runner:
```
docker model run hf.co/Agentware/d3-flash-calibrated-gguf:F32
```

Lemonade

How to use Agentware/d3-flash-calibrated-gguf with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull Agentware/d3-flash-calibrated-gguf:F32

Run and chat with the model

lemonade run user.d3-flash-calibrated-gguf-F32

List all available models

lemonade list

Hermes Agent

How to use Agentware/d3-flash-calibrated-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 Agentware/d3-flash-calibrated-gguf:F32

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 Agentware/d3-flash-calibrated-gguf:F32

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use Agentware/d3-flash-calibrated-gguf with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Agentware/d3-flash-calibrated-gguf:F32

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 "Agentware/d3-flash-calibrated-gguf:F32" \
  --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"

D3 flash calibrated GGUF derivatives

GitHub project and documentation · Usage guide · Evaluation results

Project documentation includes engine setup, evaluation methodology and artifact checksum verification.

Published 2026-10-22. See VERIFICATION.md for integrity and runtime validation status.

Community text-only derivatives of vllm-sr/d3-flash. Not an official upstream release.

Source and files

flash_all_INT8

Source: vllm-sr/d3-flash@b44536d84b31a0d7b71871eb77b9656bc7f9fdf1
File: flash-allINT8-CUDA-requantized-Q8_0-F32readout.gguf
Size: 12,515,290,176 bytes (12.515 decimal GB)
SHA-256: de009412241cdc9d683c2403d598d488c9597ad6c44b9b66a9cec17810e86bf9
Tensor types: {"F32": 178, "Q8_0": 249}
Original recipe: all_int8_g128

flash_mixed

Source: vllm-sr/d3-flash@b44536d84b31a0d7b71871eb77b9656bc7f9fdf1
File: flash-MLP-NVFP4-attention-Q8_0-CUDA-F32readout.gguf
Size: 10,099,379,392 bytes (10.099 decimal GB)
SHA-256: cfa2c90cbebf01c7ad751f7d04440f5a4dcf213b4b96a5defe9ba3437c4d9be0
Tensor types: {"F32": 274, "Q8_0": 153, "NVFP4": 96}
Original recipe: mlp_nvfp4_scaled_g16_attention_int8_g128_sfp32

Conversion and runtime

Lossy re-quantized GGUF derivatives. Earlier INT8 g128 effective weights were reconstructed through BF16 and re-encoded as Q8_0 g32. NVFP4 tensors, where present in the counts above, use the recorded native GGUF type. F32 decision readout is embedded in output.weight. This is not a lossless repack of the earlier low-bit checkpoint.

The model uses the systemone decision protocol, a systemone template and non-causal attention. Generic autoregressive chat inference is not a validated substitute. See the accompanying USAGE.md for the verified engine revision and remaining clean-install checks.

Evaluation

Variant	GPQA Diamond 196	WinoGrande 1267	GSM8K 2638	MuSR 752	CRUXEval 570
flash_all_INT8	94/196 (47.96%)	1168/1267 (92.19%)	1923/2638 (72.90%)	462/752 (61.44%)	404/570 (70.88%)
flash_mixed	99/196 (50.51%)	1162/1267 (91.71%)	1898/2638 (71.95%)	462/752 (61.44%)	397/570 (69.65%)

These are decision-index edition 0.3 task variants, not standard free-generation scores or the complete Decision Index. GSM8K has two option-count variants per original question; CRUXEval is output selection, not executable-code pass@1. Full protocol and comparative limits: EVALUATION.md. Raw benchmark questions are not redistributed here.

Limits and intended use

Research and evaluation of text-only decisions. A100 inference was recorded; L4 and CPU operation are not validated for this release. File size is not a VRAM requirement. Independent safety/performance evaluation is needed before consequential deployment. Quantization can change decisions; published accuracy does not guarantee reliability on another distribution.

License and attribution

Upstream model license: Apache-2.0. Include its license text and applicable upstream notices in the published repository. This derivative changes storage/precision and removes vision components. See LICENSE_REVIEW.md, LICENSE and NOTICE in this repository. Benchmark data and new project code have separate licensing considerations.

Project

AgentwareStudio/d3-gguf contains project documentation and the aggregate evaluation report.

Downloads last month: 212

GGUF

Model size

9B params

Architecture

qwen35

Hardware compatibility

32-bit

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for Agentware/d3-flash-calibrated-gguf

Base model

Qwen/Qwen3.5-9B-Base

Finetuned

Qwen/Qwen3.5-9B

Finetuned

vllm-sr/d3-flash

Quantized

(13)

this model