Instructions to use dendriteholdings/Ingot-27B-DFlash2 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use dendriteholdings/Ingot-27B-DFlash2 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 dendriteholdings/Ingot-27B-DFlash2:BF16
# Run inference directly in the terminal:
llama cli -hf dendriteholdings/Ingot-27B-DFlash2:BF16

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf dendriteholdings/Ingot-27B-DFlash2:BF16
# Run inference directly in the terminal:
llama cli -hf dendriteholdings/Ingot-27B-DFlash2:BF16

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 dendriteholdings/Ingot-27B-DFlash2:BF16
# Run inference directly in the terminal:
./build/bin/llama cli -hf dendriteholdings/Ingot-27B-DFlash2:BF16

Use Docker

docker model run hf.co/dendriteholdings/Ingot-27B-DFlash2:BF16

Llama new
LM Studio
Jan

vLLM

How to use dendriteholdings/Ingot-27B-DFlash2 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "dendriteholdings/Ingot-27B-DFlash2"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "dendriteholdings/Ingot-27B-DFlash2",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/dendriteholdings/Ingot-27B-DFlash2:BF16

Ollama
How to use dendriteholdings/Ingot-27B-DFlash2 with Ollama:
```
ollama run hf.co/dendriteholdings/Ingot-27B-DFlash2:BF16
```
Unsloth Desktop

How to use dendriteholdings/Ingot-27B-DFlash2 with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf dendriteholdings/Ingot-27B-DFlash2:BF16

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": "dendriteholdings/Ingot-27B-DFlash2:BF16"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use dendriteholdings/Ingot-27B-DFlash2 with Docker Model Runner:
```
docker model run hf.co/dendriteholdings/Ingot-27B-DFlash2:BF16
```

Lemonade

How to use dendriteholdings/Ingot-27B-DFlash2 with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull dendriteholdings/Ingot-27B-DFlash2:BF16

Run and chat with the model

lemonade run user.Ingot-27B-DFlash2-BF16

List all available models

lemonade list

Hermes Agent

How to use dendriteholdings/Ingot-27B-DFlash2 with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf dendriteholdings/Ingot-27B-DFlash2:BF16

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 dendriteholdings/Ingot-27B-DFlash2:BF16

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use dendriteholdings/Ingot-27B-DFlash2 with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf dendriteholdings/Ingot-27B-DFlash2:BF16

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 "dendriteholdings/Ingot-27B-DFlash2:BF16" \
  --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"

Browse Quantizations to use this model in llama.cpp, Ollama, LM Studio, or any compatible app.

Ingot-27B-DFlash2

A DFlash 2 speculative-decoding drafter for the Ingot 27B kings, fine-tuned from z-lab/Qwen3.8-27B-DFlash2 on answers generated by Ingot-IV-27B.

The drafter only proposes tokens; the king verifies every one, so it speeds up generation without changing which model answers. It runs in the Dendrite llama.cpp fork next to a PQ2_0 king.

Files

file	type	size
`Ingot-27B-DFlash2-Q8_0.gguf`	Q8_0, recommended	2.06 GB
`Ingot-27B-DFlash2-BF16.gguf`	BF16	3.86 GB

Usage

llama-server -m Ingot-XVI-27B-PQ2_0.gguf \
  -md Ingot-27B-DFlash2-Q8_0.gguf --spec-type draft-dflash --spec-draft-n-max 5 \
  -fa on --jinja

Five drafted tokens per step is the fastest setting we measured. The drafter adds about 6 GB of VRAM at a 16k context.

Results

RTX PRO 6000 Blackwell, one stream, Dendrite llama.cpp fork.

Chat prompts held out from training, 256 greedy tokens:

target	plain tok/s	with drafter tok/s	speedup	acceptance
Ingot-IV-27B	135.5	243.9	1.80x	55.8%
Ingot-XVI-27B	135.8	245.2	1.81x	56.4%
Ingot-IV-27B, upstream z-lab drafter	138	186	1.35x	43.0%

Long coding prompts, Ingot-XVI-27B, temperature 0.6, top-p 0.95:

prompt tokens	plain tok/s	with drafter tok/s	speedup	acceptance
2k	134.7	175.2	1.30x	35.5%
16k	122.8	163.4	1.33x	32.3%
60k	98.3	138.2	1.41x	30.7%

Q8_0 is faster than BF16 at the same acceptance.

Training

Start: z-lab/Qwen3.8-27B-DFlash2. 5 layers, block size 8, target features from layers 6, 20, 34, 48 and 62.
Target: Ingot-IV-27B PQ2_0, frozen. Its embeddings and output head are shared with the drafter.
Data: 36,102 prompts from dendriteholdings/ingot_corpus, answered by the king at temperature 1.0, top-p 0.95. An evaluation-sized set of 903 prompts was held out.
Objective: DFlash block teacher forcing, 64 anchors per sequence, sequences up to 3,072 tokens, position-weighted cross-entropy with gamma 4.
Optimizer: AdamW, betas 0.9 and 0.95, weight decay 0.01, gradient clip 1.0, cosine schedule, 1,499 steps over two stages, 116M tokens, 8 GPUs.
The DFlash 2 selector is unchanged from upstream.

Limitations

Trained on sequences up to 3k tokens; acceptance holds to at least 60k in our tests.
Acceptance is lower when sampling than with greedy decoding.
Speculative decoding in llama.cpp verifies drafts in batches, so outputs can differ from plain decoding at near-tied tokens. The model is the same; the floating-point path is not.
Measured on a single stream; the gain under many concurrent requests is smaller.

Citation

@article{chen2026dflash,
  title   = {DFlash: Block Diffusion for Flash Speculative Decoding},
  author  = {Chen, Jian and Liang, Yesheng and Liu, Zhijian},
  journal = {arXiv preprint arXiv:2602.06036},
  year    = {2026},
  url     = {https://arxiv.org/abs/2602.06036}
}

Upstream drafter by z-lab, Apache 2.0.