Self-Serve Deployments
Self-serve deployments provide reserved inference capacity on Crusoe's optimized inference engine and managed infrastructure. They give you predictable performance, dedicated throughput, and scalability that grows with your workload.
When to use self-serve deployments
Self-serve deployments are best when you need:
- Predictability — Consistent, predictable inference performance over time
- Sustained traffic — High-volume processing that benefits from reduced cost-per-token as you scale
- Fine-tuned models — Deployment of your own Low-Rank Adaptation (LoRA) adapters trained through Serverless Fine-Tuning offering
- Control over scaling — Reduced risk of rate limiting and direct control over the number of replicas supporting your workload
- Pay-per-use flexibility — GPU-hour billing that lets you align cost with actual usage
Self-serve vs. serverless
If your workload is sporadic or low-volume, serverless might be a better fit. Use the following table to compare the two options across the dimensions that most affect cost and performance.
| Dimension | Self-serve | Serverless |
|---|---|---|
| Capacity | Dedicated deployment with reserved GPUs | Shared multi-tenant pool |
| Rate limits | None; throughput bounded by replica count | Shared pool limits |
| Billing | GPU-hour (time-based) | Per-token |
| Model support | Base models and LoRA adapters from Serverless Fine-Tuning | Base models only; no custom or fine-tuned models |
| Cost efficiency | Best at sustained high utilization | Best for sporadic or low-volume use |
Choose an optimization profile
Every self-serve deployment is configured with one optimization profile per model. The profile determines how the inference engine is tuned for your workload.
| Profile | Optimization | Best for |
|---|---|---|
| Responsiveness | Low latency, optimized for quick responses | Interactive applications, real-time inference, latency-sensitive workloads |
| Throughput | Cost efficiency at scale, optimized for token volume | Batch processing, high-volume workflows, cost-per-token minimization |
| Balanced | Hybrid blend of throughput and responsiveness, optimized to support moderate token volume and latency | General purpose production traffic |
Supported models
Self-serve deployments support the following base models. You can also deploy these models with LoRA adapters trained through Serverless Fine-Tuning.
| Lab | Model | Input modalities | Output modalities |
|---|---|---|---|
| DeepSeek | DeepSeek V4 Flash | Text | Text |
| Gemma 4 31B IT | Image, Text | Text | |
| Meta | Llama 3.1 8B Instruct | Text | Text |
| Meta | Llama 3.3 70B Instruct | Text | Text |
| Moonshot AI | Kimi K2.6 | Text, Image | Text |
| OpenAI | GPT-OSS 20B | Text | Text |
| OpenAI | GPT-OSS 120B | Text | Text |
| Qwen | Qwen3 8B | Text | Text |
| Qwen | Qwen3 235B A22B Instruct | Text | Text |
| Qwen | Qwen3.5 2B | Image, Text | Text |
| Qwen | Qwen3.5 9B | Image, Text | Text |
| Qwen | Qwen3.6 27B | Image, Text | Text |
| Qwen | Qwen3.6 35B | Image, Text | Text |
| Z AI | GLM 5.1 | Text | Text |
Next steps
- Set up your first deployment
- For additional optimization, Contact us about Tailored Deployments
- Learn more about Managed AI