Modal
Listed at https://modal.com
Overall Rank #28 ⭐ Consider
❌ Proxy required (infrastructure on AWS/GCP overseas regions; direct CN access unreliable) | 🌍 International
💰 Token Pricing
| Type | Price | Note |
|---|---|---|
| Input | 按 GPU 秒计费：H100 $0.001097/s、A100 80GB $0.000694/s、T4 $0.000164/s | per million tokens |
| Output | 同输入（按 GPU 时间，无 idle 费用） | per million tokens |
| Cache Read | 无 prompt cache 概念（按 GPU 实际运行时间计算） | Discounted |
💡 Free Credits:
🤖 Supported Models (12)
B300 / B200 / H200 / H100 / A100 / L40S / A10 / L4 / T4 (GPU)Self-hosted LLM inference (vLLM, SGLang, TensorRT-LLM)Custom PyTorch / TensorFlow / JAX modelsStable Diffusion XL / FLUX / SD3 image generationWhisper transcription / XTTS speech synthesisEmbeddings (BGE, GTE, sentence-transformers)Sandbox (arbitrary Linux containers)
✨ Pros
- ✓Python-native serverless GPU platform with excellent developer experience
- ✓Per-GPU-second billing with zero idle fees — best cost profile for spiky workloads
- ✓Built-in Sandbox, distributed queue, distributed dict; supports inference/training/batch
- ✓Starter plan includes $30/month free credits — indie-developer friendly
- ✓One-line deploy of vLLM / SGLang / TensorRT-LLM open-source LLMs
⚠️ Cons
- ×Not a model API provider — self-hosted OSS models require selection + maintenance
- ×Cold start 1-3s latency — not suited for ultra-low-latency realtime inference
- ×Region selection surcharge 1.5-1.75x — non-primary regions are expensive
- ×No prompt-cache concept — long-context workloads cost more than Anthropic/OpenAI
🎯 Best For
Self-host open-source LLMs (Llama/Qwen/DeepSeek); spiky workloads; ML training and batch inference