Instructions to use MachineFi/Trio-Spark-DI with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use MachineFi/Trio-Spark-DI with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("image-text-to-text", model="MachineFi/Trio-Spark-DI")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("MachineFi/Trio-Spark-DI")
model = AutoModelForMultimodalLM.from_pretrained("MachineFi/Trio-Spark-DI", device_map="auto")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

llama.cpp

How to use MachineFi/Trio-Spark-DI 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 MachineFi/Trio-Spark-DI:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf MachineFi/Trio-Spark-DI:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf MachineFi/Trio-Spark-DI:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf MachineFi/Trio-Spark-DI:Q4_K_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 MachineFi/Trio-Spark-DI:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama cli -hf MachineFi/Trio-Spark-DI:Q4_K_M

Use Docker

docker model run hf.co/MachineFi/Trio-Spark-DI:Q4_K_M

Llama new
LM Studio
Jan

vLLM

How to use MachineFi/Trio-Spark-DI with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "MachineFi/Trio-Spark-DI"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "MachineFi/Trio-Spark-DI",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker

docker model run hf.co/MachineFi/Trio-Spark-DI:Q4_K_M

SGLang

How to use MachineFi/Trio-Spark-DI with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "MachineFi/Trio-Spark-DI" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "MachineFi/Trio-Spark-DI",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "MachineFi/Trio-Spark-DI" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "MachineFi/Trio-Spark-DI",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Ollama
How to use MachineFi/Trio-Spark-DI with Ollama:
```
ollama run hf.co/MachineFi/Trio-Spark-DI:Q4_K_M
```
Unsloth Desktop

How to use MachineFi/Trio-Spark-DI with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf MachineFi/Trio-Spark-DI:Q4_K_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": "MachineFi/Trio-Spark-DI:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use MachineFi/Trio-Spark-DI with Docker Model Runner:
```
docker model run hf.co/MachineFi/Trio-Spark-DI:Q4_K_M
```

Lemonade

How to use MachineFi/Trio-Spark-DI with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull MachineFi/Trio-Spark-DI:Q4_K_M

Run and chat with the model

lemonade run user.Trio-Spark-DI-Q4_K_M

List all available models

lemonade list

Hermes Agent

How to use MachineFi/Trio-Spark-DI with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf MachineFi/Trio-Spark-DI:Q4_K_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 MachineFi/Trio-Spark-DI:Q4_K_M

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use MachineFi/Trio-Spark-DI with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf MachineFi/Trio-Spark-DI:Q4_K_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 "MachineFi/Trio-Spark-DI:Q4_K_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"

Trio-Spark-DI / README.md

Raullen

Model card: vision score from the submitted Hub run

8e9c6b0 verified 14 days ago

preview code

Raw History Blame Contribute Delete

5.59 kB

---
license: cc-by-nc-4.0
base_model: Qwen/Qwen3.5-9B
base_model_relation: finetune
library_name: transformers
pipeline_tag: image-text-to-text
tags:
- decision-model
- decision-index
- vision
- gguf
extra_gated_heading: "Access Trio-Spark DI (research, non-commercial)"
extra_gated_prompt: "These weights are released under CC BY-NC 4.0 for research and evaluation only. Commercial use requires a separate license from MachineFi (or use the hosted Trio-Spark API)."
extra_gated_fields:
  Name: text
  Organization: text
  Intended use: text
  I agree to use these weights for non-commercial purposes only: checkbox
---
# Trio-Spark DI (research checkpoint)

Trio-Spark DI is a research checkpoint of MachineFi's Trio-Spark decision model, trained for the
[Decision Index](https://github.com/apolinario/decision-index) typed-decision protocol:
given a `state` and typed questions with explicit options, it returns a probability for every option.

- **Base model:** [Qwen/Qwen3.5-9B](https://huggingface.co/Qwen/Qwen3.5-9B) (Apache-2.0).
- **Weights:** `model-0000{1..4}-of-00004.safetensors` (bf16, with the base model's vision tower) load in Transformers and vLLM as `Qwen3_5ForConditionalGeneration`; these are the weights evaluated on the vision board. `trio-spark-di-q8_0.gguf` (Q8_0, about 9.5 GB, sha256 `42edc59b…`) is the same checkpoint (DI14), text only, as evaluated on the text board; `trio-spark-di-q4_k_m.gguf` (Q4_K_M, about 5.6 GB) is the same at 4 bits and scores lower.
- **Text and images.** The bf16 weights take images through the base model's vision tower; the GGUF files are text only. The hosted Trio-Spark API (v1.3) is a different, production model; this checkpoint is not what the API serves.
## Decision Index 0.3 (public suite, self-run)
| | Decision Index | Knowledge | Language | Retrieval | Tools | Arts |
|---|---:|---:|---:|---:|---:|---:|
| **Trio-Spark DI (DI14, Q8_0), full public suite** | **56.74** | 40.4 | 60.0 | 59.4 | 79.3 | 43.9 |
| Trio-Spark DI (DI14, Q4_K_M), full public suite | 55.50 | 39.2 | 59.4 | 58.7 | 76.4 | 42.7 |
All 140,178 requests of the public 0.3 suite, run with the official kit against the server in `serve/` (18 requests longer than 8,192 tokens refused, none truncated). The server is deterministic: the same request always gets the same answer.
Submitted to the Decision Index leaderboard; the official score also includes private test sets and may differ.
## Decision Index vision (public suite, self-run)
| | Vision index | CV-Bench | BLINK | RealWorldQA | CharXiv | InfographicVQA | Mind2Web | Winoground | KIE | Moderation | R-Bench-M | MMMU-Pro |
|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|
| **Trio-Spark DI (bf16)** | **67.07** | 76.1 | 52.3 | 68.1 | 71.7 | 93.4 | 70.9 | 80.2 | 92.3 | 27.1 | 25.5 | 35.7 |
All 12,210 rows of the public vision suite, run with the kit's reference `vllm-vl` engine (images up to 1.6 MP). Submitted to the Decision Index vision leaderboard; the official score also includes private test sets and may differ.
## Earlier text checkpoints
Earlier checkpoints (Q4_K_M), on a 12,001-request stratified sample of the same suite:
| | Decision Index | Knowledge | Language | Retrieval | Tools | Arts |
|---|---:|---:|---:|---:|---:|---:|
| Trio-Spark DI (DI14) | 54.47 | 38.1 | 55.8 | 58.4 | 77.3 | 43.8 |
| Trio-Spark DI (DI13, tag `di13`) | 54.08 | 38.6 | 54.5 | 58.2 | 76.6 | 43.7 |
## Running it
Trio-Spark DI is not a chat model. It answers through its decision prompt and reads a probability for every option from the model's next-letter scores, so plain chat prompts will not reproduce the scores above.
**Images (bf16 weights).** Use vLLM or Transformers; the chat template has thinking off by default (`enable_thinking=true` turns it on). The vision score above comes from the Decision Index kit's reference engine:

```sh
python -m decision_index vision pipeline --dir work/vision --engine vllm-vl --option model=MachineFi/Trio-Spark-DI --out runs/trio-spark-di-vision
```

**Text (GGUF).** The `/v1/systemone` server used for the text evaluation is in `serve/`: llama.cpp (release b11249, CUDA 12.8) with one slot per server and a small Python front end (standard library only) that evaluates the questions of a request in order and clears the slot between requests, so answers do not depend on what ran before. It is compatible with the Decision Index kit's `http` engine.

```sh
hf download MachineFi/Trio-Spark-DI --include "trio-spark-di-q8_0.gguf" "serve/*" --local-dir trio-spark-di
bash trio-spark-di/serve/setup.sh $PWD/trio-spark-di/trio-spark-di-q8_0.gguf
# then, from the Decision Index kit:
python -m decision_index pipeline --engine http --option base_url=http://127.0.0.1:8800 --option model=trio-spark --out runs/trio-spark-di
```

Text server limits: up to 8,192 tokens per decision prompt (longer requests are refused, never truncated); questions with more than 8 options are decided by an 8-way tournament (groups of 8, then a final over the group winners), so no option is dropped.

## License

The weights are released under [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/): free for research and evaluation with attribution,
**no commercial use**. For commercial use, use the [hosted Trio-Spark API](https://platform.machinefi.com/spark) or contact MachineFi.
The base model Qwen3.5-9B is Apache-2.0; its license and notice apply to the base weights (see `NOTICE`). Full license text: `LICENSE`.

The hosted, production model is documented at [MachineFi/Trio-Spark](https://huggingface.co/MachineFi/Trio-Spark).

	---
	license: cc-by-nc-4.0
	base_model: Qwen/Qwen3.5-9B
	base_model_relation: finetune
	library_name: transformers
	pipeline_tag: image-text-to-text
	tags:
	- decision-model
	- decision-index
	- vision
	- gguf
	extra_gated_heading: "Access Trio-Spark DI (research, non-commercial)"
	extra_gated_prompt: "These weights are released under CC BY-NC 4.0 for research and evaluation only. Commercial use requires a separate license from MachineFi (or use the hosted Trio-Spark API)."
	extra_gated_fields:
	Name: text
	Organization: text
	Intended use: text
	I agree to use these weights for non-commercial purposes only: checkbox
	---

	# Trio-Spark DI (research checkpoint)

	Trio-Spark DI is a research checkpoint of MachineFi's Trio-Spark decision model, trained for the
	[Decision Index](https://github.com/apolinario/decision-index) typed-decision protocol:
	given a `state` and typed questions with explicit options, it returns a probability for every option.

	- Base model: [Qwen/Qwen3.5-9B](https://huggingface.co/Qwen/Qwen3.5-9B) (Apache-2.0).
	- Weights: `model-0000{1..4}-of-00004.safetensors` (bf16, with the base model's vision tower) load in Transformers and vLLM as `Qwen3_5ForConditionalGeneration`; these are the weights evaluated on the vision board. `trio-spark-di-q8_0.gguf` (Q8_0, about 9.5 GB, sha256 `42edc59b…`) is the same checkpoint (DI14), text only, as evaluated on the text board; `trio-spark-di-q4_k_m.gguf` (Q4_K_M, about 5.6 GB) is the same at 4 bits and scores lower.
	- Text and images. The bf16 weights take images through the base model's vision tower; the GGUF files are text only. The hosted Trio-Spark API (v1.3) is a different, production model; this checkpoint is not what the API serves.

	## Decision Index 0.3 (public suite, self-run)

	\| \| Decision Index \| Knowledge \| Language \| Retrieval \| Tools \| Arts \|
	\|---\|---:\|---:\|---:\|---:\|---:\|---:\|
	\| Trio-Spark DI (DI14, Q8_0), full public suite \| 56.74 \| 40.4 \| 60.0 \| 59.4 \| 79.3 \| 43.9 \|
	\| Trio-Spark DI (DI14, Q4_K_M), full public suite \| 55.50 \| 39.2 \| 59.4 \| 58.7 \| 76.4 \| 42.7 \|

	All 140,178 requests of the public 0.3 suite, run with the official kit against the server in `serve/` (18 requests longer than 8,192 tokens refused, none truncated). The server is deterministic: the same request always gets the same answer.
	Submitted to the Decision Index leaderboard; the official score also includes private test sets and may differ.

	## Decision Index vision (public suite, self-run)

	\| \| Vision index \| CV-Bench \| BLINK \| RealWorldQA \| CharXiv \| InfographicVQA \| Mind2Web \| Winoground \| KIE \| Moderation \| R-Bench-M \| MMMU-Pro \|
	\|---\|---:\|---:\|---:\|---:\|---:\|---:\|---:\|---:\|---:\|---:\|---:\|---:\|
	\| Trio-Spark DI (bf16) \| 67.07 \| 76.1 \| 52.3 \| 68.1 \| 71.7 \| 93.4 \| 70.9 \| 80.2 \| 92.3 \| 27.1 \| 25.5 \| 35.7 \|

	All 12,210 rows of the public vision suite, run with the kit's reference `vllm-vl` engine (images up to 1.6 MP). Submitted to the Decision Index vision leaderboard; the official score also includes private test sets and may differ.

	## Earlier text checkpoints

	Earlier checkpoints (Q4_K_M), on a 12,001-request stratified sample of the same suite:

	\| \| Decision Index \| Knowledge \| Language \| Retrieval \| Tools \| Arts \|
	\|---\|---:\|---:\|---:\|---:\|---:\|---:\|
	\| Trio-Spark DI (DI14) \| 54.47 \| 38.1 \| 55.8 \| 58.4 \| 77.3 \| 43.8 \|
	\| Trio-Spark DI (DI13, tag `di13`) \| 54.08 \| 38.6 \| 54.5 \| 58.2 \| 76.6 \| 43.7 \|

	## Running it

	Trio-Spark DI is not a chat model. It answers through its decision prompt and reads a probability for every option from the model's next-letter scores, so plain chat prompts will not reproduce the scores above.

	Images (bf16 weights). Use vLLM or Transformers; the chat template has thinking off by default (`enable_thinking=true` turns it on). The vision score above comes from the Decision Index kit's reference engine:

	```sh
	python -m decision_index vision pipeline --dir work/vision --engine vllm-vl --option model=MachineFi/Trio-Spark-DI --out runs/trio-spark-di-vision
	```

	Text (GGUF). The `/v1/systemone` server used for the text evaluation is in `serve/`: llama.cpp (release b11249, CUDA 12.8) with one slot per server and a small Python front end (standard library only) that evaluates the questions of a request in order and clears the slot between requests, so answers do not depend on what ran before. It is compatible with the Decision Index kit's `http` engine.

	```sh
	hf download MachineFi/Trio-Spark-DI --include "trio-spark-di-q8_0.gguf" "serve/*" --local-dir trio-spark-di
	bash trio-spark-di/serve/setup.sh $PWD/trio-spark-di/trio-spark-di-q8_0.gguf
	# then, from the Decision Index kit:
	python -m decision_index pipeline --engine http --option base_url=http://127.0.0.1:8800 --option model=trio-spark --out runs/trio-spark-di
	```

	Text server limits: up to 8,192 tokens per decision prompt (longer requests are refused, never truncated); questions with more than 8 options are decided by an 8-way tournament (groups of 8, then a final over the group winners), so no option is dropped.

	## License

	The weights are released under [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/): free for research and evaluation with attribution,
	no commercial use. For commercial use, use the [hosted Trio-Spark API](https://platform.machinefi.com/spark) or contact MachineFi.
	The base model Qwen3.5-9B is Apache-2.0; its license and notice apply to the base weights (see `NOTICE`). Full license text: `LICENSE`.

	The hosted, production model is documented at [MachineFi/Trio-Spark](https://huggingface.co/MachineFi/Trio-Spark).