Synthetic delivers a managed inference platform that runs open‑source large language models in isolated, GDPR‑compliant datacenters with end‑to‑end encryption and no prompt or completion logging. The service provides OpenAI‑ and Anthropic‑compatible REST APIs, supports the full vLLM model catalog, and includes built‑in LoRA adapters, embedding models, and flexible flat‑rate or pay‑per‑token pricing.
Funding
Funding not disclosed
Founders
Product
Problem
Enterprises and developers often need to run open‑source large language models (LLMs) without exposing proprietary or sensitive data to external providers. Conventional hosted APIs may retain prompts and completions for model training, creating compliance and privacy risks. Additionally, integrating a wide variety of open‑source models into existing OpenAI‑compatible workflows can be technically cumbersome.
Solution
Synthetic offers a managed inference platform that executes open‑source LLMs inside isolated, GDPR‑compliant datacenters, guaranteeing that user prompts and completions are never logged or used for training. The service exposes fully OpenAI‑ and Anthropic‑compatible REST endpoints, allowing seamless drop‑in replacement for existing tooling. By leveraging the vLLM runtime, Synthetic supports any model listed in the vLLM catalog—including GLM‑4.7, Kimi K2, MiniMax M2.1, Qwen 3, and many others—without requiring custom deployment. Customers can choose a flat‑rate subscription ($30 / month) for unlimited access to always‑on models or a pay‑per‑token usage plan that bills only for actual inference. An open‑source evaluation suite, Synbad, continuously validates model behavior, delivering a 100 % pass rate on critical coding‑agent benchmarks. The platform also provides built‑in LoRA adapters, embedding models, and high‑throughput rate‑limit tiers to meet diverse workload demands.
Target Audience
The primary customers are AI researchers, software engineers, and enterprise teams that require confidential LLM inference for internal applications, data‑sensitive workloads, or compliance‑driven environments. It also serves developers building AI agents or SaaS products who need a plug‑and‑play, OpenAI‑compatible backend for a broad set of open‑source models.
Features
- Private, isolated execution environment in secure datacenters with end‑to‑end encryption and no data retention for prompts or completions.
- OpenAI‑compatible `/v1/chat/completions`, `/v1/completions`, `/v1/embeddings` and Anthropic‑compatible `/v1/messages` endpoints for drop‑in integration.
- vLLM‑driven model catalog supporting all vLLM‑compatible open‑source LLMs (e.g., GLM‑4.7, Kimi K2, MiniMax M2.1, Qwen 3, DeepSeek, Llama‑3, etc.).
- Subscription tier ($30 / month) with flat‑rate access and usage‑based tier with per‑token billing; optional $1 / day pack offering higher message limits.
- Synbad evaluation suite delivering 100 % pass rate on real‑world coding‑agent bug tests, ensuring model reliability.
- Built‑in LoRA support (rank‑8 to rank‑64, FP8 precision) and pre‑hosted embedding models for retrieval‑augmented applications.
- Rate‑limit and quota APIs (`/quotas`) for monitoring consumption and enforcing enterprise SLAs.
- SDKs and integrations for popular developer tools (Roo, Cline, Octofriend) and standard OpenAI client libraries.