Sail Research

Purpose

Sail Research is an inference API platform for open-source large models, aiming to provide high-quality, large-scale, scalable model inference services at the lowest possible cost. Its core philosophy is "pursuing token abundance," serving agents, long-running tasks, and high-throughput workloads by carefully selecting the best open-source models and optimizing to the extreme across the entire latency-cost curve.

Main Capabilities

1. Extreme Cost-Effectiveness: Pay-per-token with highly competitive pricing. For DeepSeek V4.1 Flash, real-time prices are 3.3x lower than other providers, and asynchronous prices are 6x lower. A $5 free credit is granted monthly under the Free plan and can be automatically refreshed.
2. Completion Window: Allows specifying latency tolerance in each request, giving the platform greater scheduling flexibility and thereby greatly reducing token costs at the same level of intelligence. It supports different windows such as balanced and flex, and the pricing page lists the windows supported by each model.
3. API Compatibility: Compatible with OpenAI and Anthropic's Responses, Chat Completions, and Messages APIs. Simply replace the base URL and key to migrate from an existing client within minutes.
4. Fine-Tuning and RL Rollouts: Supports self-deployed LoRA fine-tuning and can run RL rollouts on Sail's sampling path. It integrates directly with Tinker: LoRAs trained with Tinker can be sampled directly from Tinker checkpoints without uploading, and Sail can also be used as Tinker's TokenCompleter.
5. Massive Throughput and Elastic Scaling: Serves trillions of tokens per week and can handle workloads of millions, billions, or even trillions of tokens without sales conversations. The platform is designed to absorb large traffic bursts and can scale on demand.
6. Sailboxes: Provides full virtual machines (VMs), giving long-horizon agents compute that can run indefinitely with excellent economics. Paired with Sail inference, agents can read, write, build, and test continuously for hours, days, or longer without losing state.
7. Security and Compliance: Zero data retention (ZDR) by default; inference request and response data are only usedServices are not used to train models without written consent. SOC 2 Type I is already compliant, and SOC 2 Type II is scheduled to take effect on August 25, 2026. Pro and Enterprise customers may request that inference requests be retained within the United States (a surcharge applies as scheduled). Enterprise customers may sign a BAA to support HIPAA-compliant workloads, and may sign an MSA and DPA.
8. Deep optimization of the technology stack: from writing CUDA code to pursue the ultimate GPU performance, to going deep inside inference engines such as SGLang to improve efficiency, to distributing work across providers to improve robustness and utilization; using spot compute when available, and safely failing over to more reliable compute when it is not.

How to use

• An API key can be generated within seconds after registration.
• Point any OpenAI- or Anthropic-compatible client to Sail's base URL and start sending requests.
• Official SDKs are provided, and any compatible client can also be used; just replace the base URL and key. Responses, Chat Completions, and Messages all work.
• Each request can optionally specify a completion window to balance latency and cost.
• If you want to use a self-trained LoRA, you can upload the adapter to a supported model; if you train with Tinker, you can sample directly from a Tinker checkpoint and use it as a TokenCompleter.
• Costs can be estimated with the agent cost calculator, depending on the number of tokens, the chosen model, and the completion window.

Use cases

• Developers building long-horizon, background-running agents who need large-scale, low-cost inference and high throughput.
• Research teams that need to run RL rollouts or deploy self-trained LoRAs.
• Cost-sensitive production workloads that want to reduce token overhead at the same level of intelligence.
• Services with bursty, high-traffic demand that want to scale elastically without sales involvement.
• Enterprise customers that need data residency and compliance assurances (such as HIPAA, SOC 2).

Limitations and notes

• Access may be limited during periods of high demand, but the platform is designed to absorb large traffic bursts, especially with the flex completion window. Pro customers get priority model access, and Enterprise customers get the highest priority.
• U.S. inference request retention is an option for Pro and Enterprise customers and is subject to a scheduled surcharge.
• For the full list of specific models, the completion windows supported by each model, and pricing details, refer to the official pricing page and documentation.
• The materials currently do not provide the duration of free credit usage, specific prices for each plan, or SLA details, so these are not yet available.