What it is
The first step of self-hosting an LLM often gets stuck on downloading the weights: huge files, restricted networks, dropped connections, several models in line. dlmodel makes this a reliable, visible and resumable local model hub, and hands finished models straight to vLLM or Ollama.
Screenshots


The screenshots use sample tasks.
How it works
- Downloader: source auto → ModelScope ⇁ HF mirror; within each source aria2 → HTTP Range → SDK.
- Queue: an asyncio worker pool with the states queued → running → completed / failed / cancelled; pause and resume can interleave at any moment and only one of them takes effect.
- Storage: the vLLM / HF layout goes to
{MODEL_ROOT}/hf/<org>/<repo>/and Ollama to{MODEL_ROOT}/ollama, shared with the inference services.
Use cases
GPU servers on private networks
Connect to hf-mirror.com and ModelScope without a proxy, resume after disconnects, and get the full weights in one go
Preparing many models
Queue several models and versions, adjust concurrency while running, and get a group message when they finish
Model hub for inference
Share the model volume with vLLM / Ollama; the model library shows disk use, copies paths and generates start commands
Deployment
git clone https://github.com/sreyun/dlmodel.git && cd dlmodel
cp .env.example .env # set a long random ADMIN_TOKEN
docker compose up -d # open http://SERVER_IP:8080 and sign in with ADMIN_TOKEN
docker compose --profile ollama up -d # optional: start Ollama too
docker compose --profile vllm up -d # optional: start vLLM (needs the NVIDIA Container Toolkit)
The default images are served through Huawei Cloud SWR, so they can be pulled in mainland China without a proxy.
Open-source license
dlmodel is open source under the MIT license. For offline delivery on private networks or help deploying inference services, see pricing.