Instructions to use RecursiveSI/flint with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use RecursiveSI/flint with Transformers:

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

pipe = pipeline("text-generation", model="RecursiveSI/flint")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("RecursiveSI/flint")
model = AutoModelForCausalLM.from_pretrained("RecursiveSI/flint", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.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(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use RecursiveSI/flint with vLLM:

Install from pip and serve model

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

Use Docker

docker model run hf.co/RecursiveSI/flint

SGLang

How to use RecursiveSI/flint 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 "RecursiveSI/flint" \
    --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": "RecursiveSI/flint",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

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 "RecursiveSI/flint" \
        --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": "RecursiveSI/flint",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use RecursiveSI/flint with Docker Model Runner:
```
docker model run hf.co/RecursiveSI/flint
```

Research question: modular refactoring and ultra-sparse MoE experiments for K2-Horizon-MoVA

#20

by dhammagi - opened 2 days ago

Discussion

dhammagi

2 days ago

Title: Research question: modular refactoring and ultra-sparse MoE experiments for K2-Horizon-MoVA

We are exploring a research-oriented refactoring of K2-Horizon-MoVA-36B-A4B. The objective is to preserve the original model as a reproducible baseline, then investigate whether selected innovations from Recursive Flint can be incorporated without confusing model changes with serving-engine optimizations.

We would particularly appreciate guidance on the following:

Which source files and configuration fields should be treated as the canonical specification of the K2-Horizon MoE and MoVA architecture?
What is the precise relationship between the FFN expert routing and the MoVA value-expert routing? Which components can be modularized independently?
Is there a canonical method for calculating total and active parameters that includes the attention, dense, output-head, FFN-expert, and MoVA components consistently?
Which reference tests or known-good outputs should be preserved when restructuring the implementation?
What architectural changes would require retraining rather than checkpoint conversion, particularly if experimenting with a smaller active-parameter budget?
Are there additional technical reports, ablations, or community implementations that the maintainers recommend studying?

The intended refactoring separates learned transformations, learned coordination, and execution logistics into explicit interfaces while preserving their original mathematical behavior wherever possible.

We would value any corrections, implementation pointers, or suggestions for making the experiments scientifically reproducible.

dhammagi changed discussion status to closed 2 days ago

dhammagi changed discussion status to open 2 days ago

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment