Loading 9 AI Models to a Private HTTPS Server

I just finished a satisfying infrastructure task: deploying 9 machine learning models to a self-hosted file server and making them accessible via HTTPS with proper range request support. Here’s how it went.

The Challenge

The borisovai-admin project needed a reliable way to serve large AI models—from Whisper variants to Russian ASR solutions—without relying on external APIs or paying bandwidth fees to HuggingFace every time someone needed a model. We’re talking about 19 gigabytes of neural networks that need to be fast, resilient, and actually usable from client applications.

I started by setting up a lightweight file server, then systematically pulled models from HuggingFace using huggingface_hub. The trick was managing the downloads smartly: some models are 5+ GB, so I parallelized where possible while respecting rate limits.

What Got Deployed

The lineup includes serious tooling:

• Faster-Whisper models (base through large-v3-turbo)—for speech-to-text across accuracy/speed tradeoffs
• ruT5-ASR-large—a Russian-optimized speech recognition model, surprisingly hefty at 5.5 GB
• GigAAM variants (v2 and v3 in ONNX format)—lighter, faster inference for production
• Vosk small Russian model—the bantamweight option when you need something lean

Each model is now available at its own HTTPS endpoint: https://files.dev.borisovai.ru/public/models/{model_name}/.

The Details That Matter

Getting this right meant more than just copying files. I verified CORS headers work correctly—so browsers can fetch models directly. I tested HTTP Range requests—critical for resumable downloads and partial loads. The server reports content types properly, handles streaming, and doesn’t choke when clients request specific byte ranges.

Storage-wise, we’re using 32% of available disk (130 GB free), which gives comfortable headroom for future additions. The models cover the spectrum: from tiny Vosk (88 MB) for embedded use cases to the heavyweight ruT5 (5.5 GB) when you need Russian language sophistication.

Why This Matters

Having models hosted internally means zero API costs, predictable latency, and full control over model versions. Teams can now experiment with different Whisper sizes without vendor lock-in. The Russian ASR models become practical for real production workloads instead of expensive API calls.

This is infrastructure work—not glamorous, but it’s the kind of unsexy plumbing that makes everything else possible.

---

Eight bytes walk into a bar. The bartender asks, “Can I get you anything?” “Yeah,” reply the bytes. “Make us a double.” 😄