Instructions to use tiny-random/longcat-flash-lite-sparse with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use tiny-random/longcat-flash-lite-sparse with Transformers:

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

pipe = pipeline("text-generation", model="tiny-random/longcat-flash-lite-sparse")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import LongcatCausalLM
model = LongcatCausalLM.from_pretrained("tiny-random/longcat-flash-lite-sparse", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use tiny-random/longcat-flash-lite-sparse with vLLM:

Install from pip and serve model

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

Use Docker

docker model run hf.co/tiny-random/longcat-flash-lite-sparse

SGLang

How to use tiny-random/longcat-flash-lite-sparse 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 "tiny-random/longcat-flash-lite-sparse" \
    --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": "tiny-random/longcat-flash-lite-sparse",
		"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 "tiny-random/longcat-flash-lite-sparse" \
        --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": "tiny-random/longcat-flash-lite-sparse",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use tiny-random/longcat-flash-lite-sparse with Docker Model Runner:
```
docker model run hf.co/tiny-random/longcat-flash-lite-sparse
```

This is a randomly initialized tiny model based on meituan-longcat/LongCat-Flash-Lite-Sparse for testing and debugging.

Feature	Source	This repo
Checkpoint tensor elements	69.11B BF16, 17.26M F32	11.72M BF16, 5.76K F32
Safetensors weight files	138.29 GB (26 files)	23.53 MB (model.safetensors)
Logical layers (attention sublayers)	14 (28)	4 (8)
Hidden width / dense MLP / expert MLP	3,072 / 6,144 / 1,024	8 / 64 / 32
Q heads / MLA nope-rope-v head dims	32 / 128-64-128	8 / 64-64-128
MLA q-lora / kv-lora rank	1,536 / 512	32 / 256
Routed / identity zero experts, top-k	256 / 128, 12	32 / 128, 12
LSA index heads / head dim / top-k	16 / 128 / 2,048	Same
CLI factor / local + init tokens	2 / 1,024 + 16	Same
Over-embedding tables	12 tables, ~10.22M rows each	2 tables, ~524k rows
MTP/NextN draft layers	1 (with LSA indexer)	Same
Attention / RoPE	LSA + deepseek-yarn, 1M context	Same
Vocabulary rows / tying	131,072 / untied	Same
Routers / correction bias	FP32	Same
Stored precision	BF16	Same

The source repo ships no model code and no Transformers implementation (config.json has no model_type or auto_map); SGLang recognizes its LongcatCausalLM architecture through its longcat_flash config parser and DSA indexer. The tiny tensors were generated directly against that loader's state-dict schema and key remapping, including the nested model.ngram_embeddings.{embedders,post_projs}.* over-embedding names (newer SGLang also accepts the source checkpoint's flattened oe_embed_tokens*.weight aliases) and the model.mtp.* NextN draft. The n-gram hash modulus is internal embedding machinery rather than an externally visible task space, so it follows the smaller tiny-random/longcat-flash-lite modulus; the 131,072 task vocabulary rows and token ids are unchanged. No local Transformers execution path exists; serve with SGLang (0.5.23.dev384 or newer, which includes LongCat LSA) using the command below.

Example usage

SGLang serving (✅ verified)

LSA is implemented in SGLang (main branch; the LongCat Sparse adaptation landed via PR #32918). The checkpoint keeps the source's LongcatCausalLM architecture and serves text only, like the upstream deployment guide.

python3 -m sglang.launch_server \
  --trust-remote-code \
  --model scratch/longcat-flash-lite-sparse \
  --host 0.0.0.0 \
  --port 30000 \
  --tp-size 4 \
  --max-running-requests 64 \
  --mem-fraction-static 0.93 \
  --chunked-prefill-size 2048 \
  --dsa-prefill-backend fa3 \
  --kv-cache-dtype bfloat16

curl http://127.0.0.1:30000/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{"messages": [{"role": "user", "content": "Hello, world!"}], "max_tokens": 16}'

Creating this repo

Code: .meta/create_model.py
Model structure: .meta/model.log
Environment: .meta/env.json

Change log

2026-11-03: Initial version. Direct SGLang-schema checkpoint: Transformers has no implementation for the LongCat Sparse Attention (LSA) repo (no model_type/auto_map), so tensors follow the pinned sgl-project/sglang main loader (longcat_flash.py plus the DSA Indexer). Four dual-sublayer logical layers (eight attention sublayers) keep the shortcut-MoE topology with 32 routed and 128 weightless identity zero experts at top-k 12, compressed MLA ranks (q 32 / kv 256, the Hopper FA3 compressed-latent q_v kernel floor; the source uses kv 512) with 8 Q heads, 64/64 nope/rope and the source 128 value head dim (qk 128 + v 128 matches the Hopper FA3 MHA/MLA head-dim windows), and LSA indexers with the source 16 heads / head dim 128 on every even physical attention layer (CLI factor 2). One MTP/NextN draft layer with its own indexer is serialized under model.mtp.*; over-embeddings use the loader-native nested ngram names and the sibling tiny modulus (ratio 4, 2 x 2 windows) instead of the source 10.2-million-row hash tables.

This repo is generated by GPT-6.1-Sol.

Downloads last month: -

Safetensors

Model size

11.7M params

Tensor type

F32

BF16

Model tree for tiny-random/longcat-flash-lite-sparse

Base model

meituan-longcat/LongCat-Flash-Lite-Sparse

Finetuned

(3)

this model