1M context window
Selected open-weight models support context windows up to 1 million tokens for long documents and long-running agents.
One OpenAI-compatible API for
The open-weight inference cloud.
Enterprise-grade, OpenAI-compatible inference for the open-weight model ecosystem. One API, transparent per-token pricing, no lock-in.
Open Scale API
from openai import OpenAI
client = OpenAI(
base_url="https://api.openscalecloud.com/v1"
)
base_url="https://api.openscalecloud.com/v1"
)
response = client.chat.completions.create(
model="qwen",
stream=True
)
model="qwen",
stream=True
)
StreamingOpenAI-compatible
QwenLlamaDeepSeekMistralGPT-OSSGLMKimi / MoonshotGemmaPhiYiMistral SmallSmolLMQwenLlamaDeepSeekMistralGPT-OSSGLMKimi / MoonshotGemmaPhiYiMistral SmallSmolLM
1M
Max context window
99.9%
Uptime
1
OpenAI-compatible endpoint
The catalog
The families that matter. One endpoint.
We are onboarding the major open-weight model families behind a single OpenAI-compatible API. New models and families ship as they release.
Model catalog
01
Qwen family
02
Llama family
03
DeepSeek family
04
Mistral family
05
GPT-OSS family
06
GLM family
+ 6 more families on the roadmap, with new models added as they ship.
Why Open Scale
Built for production inference
The details that make this infrastructure ready for real production traffic.
Prefix caching built in
Automatic prefix caching at the inference layer means repeated system prompts and conversation history are billed at a fraction of the cost and return far faster.
OpenAI-compatible, drop-in
OpenAI-compatible by design. The Vercel AI SDK, LangChain, n8n, and other standard tooling work against a single endpoint, so the open ecosystem plugs in without rewrites.
Tools, reasoning & vision
First-class function calling, streamed tool calls, toggleable reasoning mode, structured outputs, and multimodal (text + image) input on the models that support it.
Global, fault-tolerant fleet
Inference is spread across a global fleet of high-performance GPUs with automatic failover. One region going down never takes your requests down with it.
Zero-data-retention path
We don't train on your prompts. A zero-data-retention, no-log serving path is available for workloads that require it.
Infrastructure
A global fleet, engineered for tokens-per-dollar
From a single request to a fleet of GPUs, the same open-weight model served reliably wherever you are.
One OpenAI-compatible endpoint
A single base URL for the full model catalog. Auth, validation, and streaming are handled at the edge of the network.
Smart routing to the fleet
Requests are routed to the best-fit healthy GPU based on model, region, and load, with warm-session affinity for cache reuse.
Hardware-efficient inference
Models run on optimized vLLM deployments across our global fleet, tuned for tokens-per-second and time-to-first-token.
Streamed back, usage included
Responses stream immediately with accurate per-token usage and cached-token accounting, so you bill what you actually use.
Reach the open-weight models that matter.
One OpenAI-compatible endpoint, reachable through the platforms you already use. No lock-in.