NebulaTele

NebulaTele is a vision-language model trained by ZTE based on Qwen3.8-27B. It supports both image and video understanding, with native thinking mode and flexible reasoning control.

Deployment

vLLM

Copy and run the following command (adjust paths and port as needed):

python3 -m vllm.entrypoints.openai.api_server \
   --host=0.0.0.0 \
   --port=8000 \
   --served-model-name=ntele-sft \
   --model=/path/to/your/checkpoint \
   --tokenizer=/path/to/your/checkpoint \
   --tensor-parallel-size=4 \
   --distributed-executor-backend=mp \
   --max-model-len=50000 \
   --tool-call-parser=qwen3_coder \
   --enable-chunked-prefill \
   --enable-prefix-caching \
   --reasoning-parser=qwen3 \
   --enable-auto-tool-choice \
   --trust-remote-code \
   --language-model-only

Note:

Replace the --model and --tokenizer paths with your own checkpoint directory.

Change --port=8000 to your desired port.

Adjust --tensor-parallel-size to match your number of GPUs.

Remove --language-model-only if you need vision capabilities.

--max-model-len can be increased up to 262144 natively (or more with RoPE scaling).

SGLang

Copy and run the following command for SGLang:

python -m sglang.launch_server \
   --model-path /path/to/your/checkpoint \
   --tokenizer-path /path/to/your/checkpoint \
   --context-length 50000 \
   --tp 4 \
   --host 0.0.0.0 \
   --port 8000 \
   --served-model-name ntele-sft

Note:

Update --model-path and --tokenizer-path with your actual checkpoint location.

Set --tp to the number of GPUs.

Change --port and --context-length as needed.

For vision support, add --enable-multimodal (if your SGLang version supports it) and remove any --language-model-only equivalent.

After starting either server, you can send requests to the OpenAI-compatible endpoint at http://<host>:<port>/v1/chat/completions.