Model Catalog
All models deploy to a dedicated T4 GPU runtime. Each endpoint is OpenAI-compatible — no SDK changes needed.
meta-llama/Llama-3-8B-InstructMeta's LLaMA 3 8B — fast, capable, OpenAI-compatible. Best balance of quality and speed for production chat APIs.
mistralai/Mistral-7B-Instruct-v0.3Mistral 7B — fast, open-weights, great for production. Slightly smaller memory footprint than LLaMA 3.
microsoft/phi-3-mini-4k-instructMicrosoft Phi-3 Mini — the most capable small model. Very low cost, ideal for high-volume workloads where speed matters more than raw capability.
mistralai/Mixtral-8x7B-InstructMixture-of-experts architecture — near GPT-4 quality on many benchmarks. Quantized to run on a single T4.
openai/whisper-large-v3OpenAI's Whisper — state-of-the-art speech recognition in 99 languages. Returns timestamped transcripts.
OpenAI compatibility
All text-generation models expose an OpenAI-compatible API at /v1/chat/completions. The model name in the request body should match the Ollama model name shown above (e.g. llama3:8b, not the full catalog ID).
Pricing
GPU deployments are billed by uptime at $0.526/hour per GPU replica. Deployments configured with min_replicas: 0 scale to zero when idle and incur no charges. Set a budget cap per deployment from the dashboard to avoid surprises.