Instructions to use TassP0lar/Spark-X2.6-4B-Recursive72 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use TassP0lar/Spark-X2.6-4B-Recursive72 with Transformers:

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

pipe = pipeline("text-generation", model="TassP0lar/Spark-X2.6-4B-Recursive72", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("TassP0lar/Spark-X2.6-4B-Recursive72", trust_remote_code=True, device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use TassP0lar/Spark-X2.6-4B-Recursive72 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "TassP0lar/Spark-X2.6-4B-Recursive72"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "TassP0lar/Spark-X2.6-4B-Recursive72",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/TassP0lar/Spark-X2.6-4B-Recursive72

SGLang

How to use TassP0lar/Spark-X2.6-4B-Recursive72 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 "TassP0lar/Spark-X2.6-4B-Recursive72" \
    --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": "TassP0lar/Spark-X2.6-4B-Recursive72",
		"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 "TassP0lar/Spark-X2.6-4B-Recursive72" \
        --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": "TassP0lar/Spark-X2.6-4B-Recursive72",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use TassP0lar/Spark-X2.6-4B-Recursive72 with Docker Model Runner:
```
docker model run hf.co/TassP0lar/Spark-X2.6-4B-Recursive72
```

Spark X2.6 4B Recursive72

Spark X2.6 4B Recursive72 (TassP0lar/Spark-X2.6-4B-Recursive72) is the 72-visit recurrent-depth (Loop72) + per-visit Paged Latent Sparse Attention (D-LSA) model built strictly atop the frozen BF16 XHToken/Spark-X2.6-4B (revision=64ca8f2a295b8ceb2757b9d4749a6558c4759e5b) 36-layer hybrid backbone.

Architecture Summary

Base Backbone: XHToken/Spark-X2.6-4B (36 physical layers = 9 full-attention + 27 sliding-window layers, native GELU, frozen BF16).
72-Visit Recurrent Schedule (4 + 24 + 24 + 12 + 8 = 72):
- Prelude L0..L3 (1 pass = 4 visits)
- Core Pass 1 L4..L27 (24 visits)
- Core Pass 2 L4..L27 (24 visits)
- Inner Pass 3 L12..L23 (12 visits)
- Coda L28..L35 (1 pass = 8 visits)
Differentiable TwoStreamMHC Controllers (10,078,176 params):
- Bounded-gain Schema v2 (gain_mode='sigmoid', mhc_schema_version=2), guaranteeing $(1 - g)I + gP \in \mathcal{B}_2$ across all 60 core visits.
- Mean Pass-2 Gain (Attn/FFN): 0.009997 / 0.010014
- Mean Pass-3 Gain (Attn/FFN): 0.004999 / 0.005004
18-Visit VisitAwareDLSAIndexBank:
- Per-visit (W_k, W_q) $\mathbb{S}^{63}$ projections distilled on the trained learned_mhc residual stream.
- Locked mhc_checkpoint_sha256: 7cad7ffc33e7962f445a037f8dddb42b5250da30d1782c8f76c4138c42057456

Quickstart

import torch
from load_inference_v03 import load_sparkx_loop72_lsa

model, tokenizer = load_sparkx_loop72_lsa(
    repo_or_dir='TassP0lar/Spark-X2.6-4B-Recursive72',
    base_repo='XHToken/Spark-X2.6-4B',
    revision='64ca8f2a295b8ceb2757b9d4749a6558c4759e5b',
    recurrence_mode='learned_mhc',
    attach_sparse_attention=True,
)
model.eval()
print(model.execution_manifest())

Downloads last month: 502

Safetensors

Model size

4B params

Tensor type

BF16

Model tree for TassP0lar/Spark-X2.6-4B-Recursive72

Base model

XHToken/Spark-X2.6-4B-Base

Finetuned

XHToken/Spark-X2.6-4B

Finetuned

(21)

this model