Introducing Self-Serve Deployments for production-scale inference

Table of contents

This is some text inside of a div block.

Inference demand is compounding. Artificial Analysis found that reasoning models use up to 20x more tokens than non-reasoning models to complete the same task. Agentic workloads don't send one request. They send a chain: planning, tool use, reasoning, verification. Reasoning models push it further, burning more compute per call than the last generation.

More calls, more compute, more cost. At this volume, cost is capability. A small per-call premium, multiplied across production traffic, turns into real budget fast. That's where open models win, and the quality gap that used to justify paying more is closing.

Open models have caught up. Now it's about closing the infrastructure gap.

This is why Crusoe Intelligence Foundry now offers three ways to run Managed Inference: Serverless Inference, Self-Serve Deployments, and Tailored Deployments, with Self-Serve Deployments new this week.

How Crusoe scales with you: three deployment options

Developers can choose Serverless Inference APIs for quick experimentation with off-the-shelf models, with no infrastructure to manage; Self-Serve Deployments to put their own base or fine-tuned open model into production; or Tailored Deployments for dedicated, custom managed inference on any fine-tuned or proprietary model with SLA-backed performance.

All three options run on Crusoe's inference engine, powered by MemoryAlloy^TM technology, a cluster-native memory fabric that persists across nodes and eliminates redundant prefill for superior price performance. On Artificial Analysis, that same engine ranked first for output speed, reaching 430+ output tokens per second on Kimi K2.6 and K2.7.

Serverless Inference

The fastest on-ramp. Access a curated library of open models (GLM 5.2, Kimi K2, Qwen 3, DeepSeek v4, Gemma 4, Llama, GPT-OSS, and more) via a fully managed, OpenAI-compatible API. Generate an API key in minutes, pay per token, and start serving. No setup, no configuration.

With the pace of AI innovation, flexibility isn't a feature; it's a requirement. Developers rely on serverless inference APIs to test the latest Kimi K2, Qwen, or DeepSeek models without provisioning infrastructure, swap models as the landscape shifts, and pay only for what they use. Crusoe Serverless Inference provides this seamless on-ramp, powered by a purpose-built engine underneath.

Self-Serve Deployments (new)

Self-Serve Deployments gives you dedicated inference endpoints with predictable costs and performance matched to your workload's actual shape. Select your model, pick an optimization profile, and deploy.

Throughput is the right profile for high-volume concurrent request pipelines where maximizing tokens processed matters most, such as a document-classification job or a large-scale summarization pipeline. Responsiveness is the right profile for user-facing products where latency is what users feel, such as a chat agent calling tools mid-conversation. Balanced is a profile option that blends throughput and responsiveness by holding a healthy per-user decode speed while keeping per-GPU throughput in consideration. It's the best starting point when your traffic mixes interactivity and high-volume requests, or when you just want a strong default.

If you've fine-tuned a model with Crusoe Serverless Fine-Tuning, deployment is one click from the same interface in Crusoe Intelligence Foundry. No exporting weights, no re-platforming, no new vendor to onboard. Your model stays in your tenancy the entire time.

Self-Serve Deployments are billed by GPU per hour. NVIDIA H100 80GB is $5.50 per hour. NVIDIA H200 141GB is $6.00 per hour. The cost for your deployment is driven by the optimization profile you choose and the number of replicas allocated. Contact sales for monthly and volume rates.

Tailored Deployments

The inference-as-a-service option. Tailored Deployments provide bespoke optimization to maximize the performance of your AI inference workload, giving you tangible outcomes without the technical overhead. Work directly with Crusoe's Engineering team for the highest level of optimization and a dedicated, benchmarked, SLA-backed endpoint. You bring the model and define the requirements. We handle the rest.

“We are firmly entering the agentic era where swarms of digital minds proactively assist humans. This requires fundamentally rethinking the entire stack: models, harnesses, and infra. Crusoe is a recognized leader in the AI inference space, helping Yutori push out the Pareto frontier of capabilities, latency, and cost.” — Dhruv Batra, Co-Founder & Chief Scientist, Yutori

Here's how the three deployment options compare

	Serverless Inference	Self-Serve Deployments	Tailored Deployments
When to use	Early-stage exploration	Reserved capacity, fine-tuned model inference	Complex or proprietary model optimization
Models	Open-source models	Your base or fine-tuned	Any, including your proprietary
Pricing	Per token	Per GPU per hour	Custom
SLA	No	No	Yes

Built for speed, scale, and control

Self-Serve Deployments reflect a deliberate conviction: that flexibility and performance optimization should not be at odds. Teams should be able to match infrastructure to their workload without sacrificing speed, scale, or budget control.

That's what Crusoe Managed Inference is built on: the belief that teams should be able to build AI ambitiously, at whatever stage they're in, without infrastructure becoming the constraint.

Get started

Self-Serve Deployments is available now in Crusoe Intelligence Foundry. Deploy your model.