POST /v1/chat/completions
142 tok/s
Illustrative interface — your model, your endpoint, your data.
Choose from curated open models or point at any vLLM-supported repository on Hugging Face — instruct, chat, code, embedding, or vision-language.
Select a GPU flavor and starting size from your WECORE GPU capacity. Preflight checks confirm VRAM and quota fit the model before it launches.
WECORE deploys vLLM and a chat web app for you — model weights pulled, served, and health-checked, exposed behind your own per-service API key.
Use the built-in web chat, or point your apps at the OpenAI-compatible endpoint. Watch tokens/sec, latency, and GPU use on the built-in dashboards.
Stand up an internal assistant or a retrieval-augmented chatbot over your own documents — a private alternative to public chat services, with sources that stay inside your network.
Serve a code model behind the OpenAI-compatible API to power IDE assistants, code review bots, and tool-using agents — billed as GPU time, not per-token to a vendor.
Deploy a vision-language model for image understanding, screenshot Q&A, and document extraction with the image-analysis interface, all on accelerated nodes.
model: meta-llama/Llama-3.1-8B-Instruct
backend: vllm (openai-compatible)
gpu: 1 × A100-80GB
ui: web chat + image analysis
expose: https + per-service API key
status: SERVING · 142 tok/s
Start from vetted open models with sane defaults, or bring any vLLM-supported Hugging Face repository — gated weights pulled with your own token.
A stable /v1 chat and completions API means existing SDKs, agents, and tools point at WECORE by changing one base URL.
Each model runs as a managed service on a dedicated server, and the fit — GPU memory, context bounds, plan eligibility — is checked before anything is created.
Ready dashboards surface throughput, latency, queue depth, and GPU utilization from day one.
Prompts, completions, and weights stay on operator storage inside your cloud — no third-party API calls, no data egress to a model vendor.
Estimate VRAM, GPU hours, and wallet impact before a model launches, and stop paying the moment you stop an idle model.
It is a self-hosted LLM launcher: you choose an open model, and WECORE deploys it on your GPU nodes with a chat web interface and an OpenAI-compatible API. The pattern is inspired by StackHPC's Azimuth LLM apps, adapted to run on WECORE GPU infrastructure.
Any model supported by vLLM served from Hugging Face — including Llama, Qwen, Mistral, DeepSeek, Gemma, and vision-language models. We ship curated presets with sensible defaults, and you can point at your own repository; gated weights are pulled with your own access token.
Through an OpenAI-compatible /v1 chat and completions API. Existing SDKs, agent frameworks, and tools work by changing only the base URL — no rewrite. A built-in web chat (and an image-analysis interface for vision models) is included for people who don't want to write code.
Nowhere outside your cloud. Prompts, completions, model weights, and logs stay on operator storage inside your WECORE infrastructure. There are no third-party API calls and no data egress to a model vendor, which makes it a strong fit for sensitive and regulated workloads.
On GPU capacity inside your own WECORE cloud — never on shared vendor hardware. Each model runs as a managed service on a dedicated server, and the endpoint is protected by a per-service API key. You own the GPUs; WECORE runs the serving layer on them.
Yes. Ready dashboards expose tokens/sec, latency, queue depth, and GPU utilization. Because you run on GPUs you already pay for, cost is GPU time — and stopping an idle model stops the cost with it.
Those are general-purpose platforms you build on: KaaS for containerized applications, the HPC platform for Slurm-based batch compute. LLMaaS is a finished product rather than a platform — it serves open models behind a chat UI and an OpenAI-compatible API in one step, so you never assemble the inference stack yourself. You don't need either product to use it.