Instructions to use dendriteholdings/Ingot-XIX-27B-Q2 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-XIX-27B-Q2 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-XIX-27B-Q2:Q2_0
# Run inference directly in the terminal:
llama cli -hf dendriteholdings/Ingot-XIX-27B-Q2:Q2_0

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf dendriteholdings/Ingot-XIX-27B-Q2:Q2_0
# Run inference directly in the terminal:
llama cli -hf dendriteholdings/Ingot-XIX-27B-Q2:Q2_0

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-XIX-27B-Q2:Q2_0
# Run inference directly in the terminal:
./build/bin/llama cli -hf dendriteholdings/Ingot-XIX-27B-Q2:Q2_0

Use Docker

docker model run hf.co/dendriteholdings/Ingot-XIX-27B-Q2:Q2_0

Llama new
LM Studio
Jan

vLLM

How to use dendriteholdings/Ingot-XIX-27B-Q2 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "dendriteholdings/Ingot-XIX-27B-Q2"
# 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-XIX-27B-Q2",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/dendriteholdings/Ingot-XIX-27B-Q2:Q2_0

Ollama
How to use dendriteholdings/Ingot-XIX-27B-Q2 with Ollama:
```
ollama run hf.co/dendriteholdings/Ingot-XIX-27B-Q2:Q2_0
```
Unsloth Desktop

How to use dendriteholdings/Ingot-XIX-27B-Q2 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-XIX-27B-Q2:Q2_0

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-XIX-27B-Q2:Q2_0"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use dendriteholdings/Ingot-XIX-27B-Q2 with Docker Model Runner:
```
docker model run hf.co/dendriteholdings/Ingot-XIX-27B-Q2:Q2_0
```

Lemonade

How to use dendriteholdings/Ingot-XIX-27B-Q2 with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull dendriteholdings/Ingot-XIX-27B-Q2:Q2_0

Run and chat with the model

lemonade run user.Ingot-XIX-27B-Q2-Q2_0

List all available models

lemonade list

Hermes Agent

How to use dendriteholdings/Ingot-XIX-27B-Q2 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-XIX-27B-Q2:Q2_0

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-XIX-27B-Q2:Q2_0

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use dendriteholdings/Ingot-XIX-27B-Q2 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-XIX-27B-Q2:Q2_0

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-XIX-27B-Q2:Q2_0" \
  --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"

Ingot-XIX-27B-Q2

A public mirror of a model that won on Ingot, a Bittensor subnet where miners quantize a fixed teacher model and the validator crowns the copy that stays closest to it.

It is Prism ML's Ternary Bonsai 2 27B, refined: the same ternary codes, with the rest of its weights re-tuned by the miner to bring it closer to Qwen3.8-27B.

Submitted by miner hotkey 5HgX5p6KZVij3qMQNRELKffP27TqXf99srbii5xK3gyqZUdB (uid 49)
Crowned from submission de088987-bcbb-433c-be11-6077280d42c3

The weights are the miner's upload, byte for byte (Ingot-XIX-27B-PQ2_0.gguf); Ingot-XIX-27B-PTQ1_0.gguf repacks them losslessly at 1.75 bits per weight. Credit belongs to the miner, and to Prism ML for Bonsai 2.

Files

file	bits per weight	size
`Ingot-XIX-27B-PTQ1_0.gguf`	1.75	5.95 GB
`Ingot-XIX-27B-PQ2_0.gguf`	2.125 (the same ternary weights in a 2-bit container)	7.21 GB

Both files hold the same ternary weights. Ingot-XIX-27B-PTQ1_0.gguf is Ingot's lossless repack of the miner's file into Prism's PTQ1_0 format: every 128-weight block keeps its codes and fp16 scale, and every weight decodes to the same value. The same repack of Bonsai 2's PQ2_0 file reproduces Prism's published PTQ1_0 file byte for byte. The 1.75 bpw file is smaller; the 2-bit file decodes faster on GPUs.

How close it is to the teacher

Per answer token, against the BF16 teacher reading the same answer: KL(teacher ‖ model) of the next-token distributions, and the hidden-state distance |h − h_teacher| / |h_teacher| averaged over every 8th layer. Lower is closer.

	KL	hidden-state distance
this model, in its winning eval	0.09246	0.3173
the king it beat, same eval	0.09418	0.3201

Speed

Decode tokens per second at batch 1, measured on Ingot king IV with the Dendrite fork at the commit above. Every q2 king has the same file sizes and layout, so these numbers hold for all of them within run-to-run noise.

GPU

llama-bench, 128 generated tokens (tg128) and a 512-token prompt (pp512), one GPU, -ngl 99 -fa 1.

GPU	file	decode tok/s	prompt tok/s
NVIDIA RTX PRO 6000 Blackwell (server edition, 96 GB)	`Ingot-XIX-27B-PQ2_0.gguf`	137	3,865
NVIDIA RTX PRO 6000 Blackwell (server edition, 96 GB)	`Ingot-XIX-27B-PTQ1_0.gguf`	145	3,865

Other GPUs: PrismML reports 142 tok/s for Bonsai 2 PQ2_0 on an RTX 5090; we have not measured other cards.

Mac

Mac Studio M3 Ultra (80-core GPU, 512 GB), llama-bench, tg128 and pp512, Metal. Two builds: PrismML's fork as published, and the Dendrite fork whose Metal patch fuses the Hadamard rotation and the linear-attention state update for this model family.

build	file	decode tok/s	prompt tok/s
PrismML llama.cpp fork, stock	`Ingot-XIX-27B-PQ2_0.gguf`	44.5	432
PrismML llama.cpp fork, stock	`Ingot-XIX-27B-PTQ1_0.gguf`	43.1	368
Dendrite fork, Metal patch	`Ingot-XIX-27B-PQ2_0.gguf`	56.4	432
Dendrite fork, Metal patch	`Ingot-XIX-27B-PTQ1_0.gguf`	52.8	368

Running it

Ingot-XIX-27B-PQ2_0.gguf is a gguf with the layout, tokenizer and Hadamard rotation of dendriteholdings/Ingot-genesis-27B-Q2 (Ingot-genesis-27B-Q2.gguf), its ternary codes Bonsai 2's and the rest of its weights the miner's. Stock llama.cpp does not run it; the Dendrite fork of PrismML's llama.cpp does, at the commit Ingot scored it with, c96baf4:

git clone https://github.com/DendriteLabs/llama.cpp && cd llama.cpp && git checkout c96baf4abe24d11a50f326e40b6223b687449d63
cmake -B build -DGGML_CUDA=ON && cmake --build build -j
build/bin/llama-server -m Ingot-XIX-27B-PQ2_0.gguf -ngl 99 -fa on --jinja

Either file runs the same way: pass -m Ingot-XIX-27B-PTQ1_0.gguf for the 1.75 bpw one.

Faster generation

We trained a DFlash 2 speculative-decoding drafter for this model family, dendriteholdings/Ingot-27B-DFlash2. It proposes several tokens at a time and the model verifies each one, so it decodes about 1.8x faster in greedy single-stream chat and about 1.3-1.4x when sampling, at the cost of about 6 GB more VRAM. The same fork runs it:

hf download dendriteholdings/Ingot-27B-DFlash2 Ingot-27B-DFlash2-Q8_0.gguf --local-dir .
build/bin/llama-server -m Ingot-XIX-27B-PQ2_0.gguf -md Ingot-27B-DFlash2-Q8_0.gguf --spec-type draft-dflash --spec-draft-n-max 5 -ngl 99 -fa on --jinja

With images

The file is the text model. For image input, add the vision tower of prism-ml/Ternary-Bonsai-2-27B-gguf as a projector (Ternary-Bonsai-2-27B-mmproj-Q8_0.gguf, Ternary-Bonsai-2-27B-mmproj-BF16.gguf). It was made for genesis, not for this model's weights, and Ingot scores the text model only, so image quality is untested:

hf download prism-ml/Ternary-Bonsai-2-27B-gguf Ternary-Bonsai-2-27B-mmproj-Q8_0.gguf --revision b072e1d3b35a0a630cece372c2127528e0994386 --local-dir .
build/bin/llama-server -m Ingot-XIX-27B-PQ2_0.gguf --mmproj Ternary-Bonsai-2-27B-mmproj-Q8_0.gguf -ngl 99 -fa on --jinja

The same weights as an MLX pack for Apple Silicon: dendriteholdings/Ingot-XIX-27B-MLX-2bit.

Downloads last month: 225

GGUF

Model size

27B params

Architecture

qwen35

Hardware compatibility

1-bit

2-bit

Model tree for dendriteholdings/Ingot-XIX-27B-Q2

Base model

Qwen/Qwen3.8-27B

Quantized

prism-ml/Ternary-Bonsai-2-27B-gguf

Finetuned

(27)

this model