The managed inference platform for open models.

Companies running AI in production who want off the closed-API treadmill, from Series A up to enterprises. If you're moving workloads onto open models for cost, control, or compliance reasons, and you'd rather not build and operate an inference stack yourself, we built this for you.

Open-weight frontier models: GLM 5.2, DeepSeek V4 Flash, ThinkingCap Qwen3.6 27B, Qwen3.6 35B (A3B), Qwen3-VL 30B (vision), and Kimi K3 today, with DeepSeek V4 Pro, Gemma 4, OpenAI GPT-OSS, NVIDIA Nemotron, and Qwen3.8 rolling out next, plus your own fine-tunes and distilled models. Anything that runs on vLLM or SGLang works out of the box, and we stand up custom runtimes for models that don't. Your models get the same API, infrastructure, and audit trail as catalog models; custom model hosting is arranged as part of an enterprise agreement.

Yes, that's the primary case, and it's what the platform is tuned for. Most traffic runs on realtime sync endpoints (/v1/chat/completions, /v1/messages, blocking /v1/responses), and independent measurement backs the performance: in Requesty's live gateway comparison (mid-August 2026), sference has the highest throughput of any provider on all three models we serve there (Kimi K3 at 76 tok/s, DeepSeek V4 Flash at 157 tok/s, GLM 5.2 at 211 tok/s), and the lowest average latency on Kimi K3 (426 ms) and DeepSeek V4 Flash (861 ms). Dedicated deployments give latency-sensitive workloads reserved capacity on top of that.

No. That trade-off is usually assumed rather than measured, so we point at third-party numbers instead of our own: on Requesty's gateway, where every provider is compared on live traffic, the European route is the fastest one on every model we serve there. Every request is processed on European GPUs, outside US CLOUD Act jurisdiction, architecturally rather than by policy, and that costs you nothing in latency or throughput.

The API is OpenAI-compatible, so most clients switch with a base-URL change. Before you switch anything, we benchmark your real prompts across candidate open models side by side against your current provider, and our engineers work with your team through prompt adjustments and integration. You migrate on evidence, not on a leaderboard.

Serverless is shared endpoints billed per token, with no minimums, so you start immediately. Realtime by default, with flex (discounted, lower-priority sync via service_tier) and a 24h async window available for batch and background work. Dedicated is GPU capacity reserved for your models: predictable latency, guaranteed throughput, billed per GPU-hour on monthly commitments. Most teams start serverless and move hot paths to dedicated as volume grows; enterprise agreements add committed-spend discounts across both.

Running vLLM or SGLang as standalone servers leaves GPUs idle between requests, adds a network hop to every token, and forces server-per-model capacity planning. Our scheduler drives those same engines as in-process libraries, feeding them from priority queues with backpressure, so interactive requests are never stuck behind background work, latency stays low, and utilization stays high around the clock. You get the performance and the efficiency without building or operating any of it.

Every model is addressed by an exact version string rather than a moving "latest" alias, so the version you validated is the version serving your traffic. Closed-API providers move the model under a stable name and prompts regress without warning; here the version is explicit in the request and recorded on every response, so you can always tell which checkpoint served a given call. Before moving to a newer version, rerun your evals against it and switch when you are satisfied.

Every request is processed on European GPUs, outside US CLOUD Act jurisdiction, guaranteed architecturally rather than by policy. We never train on your data, contractually and technically. Inputs are deleted per your configured retention policy; compliance logs (model version, region, timestamps) are kept for your audit trail. GDPR, DORA, and EU AI Act ready, DPA included.

Create an account at app.sference.com and make your first API call on serverless, with no minimum spend. If you're evaluating a migration or need dedicated capacity, reach out and bring one workload with a sample of real traffic; we'll come back with benchmark numbers and pricing for your volume.