Superlinked

Superlinked Serverless Inference Engine (SIE) is Apache 2.0 software for open-model inference in a customer-controlled Kubernetes cluster. It supports AWS, Google Cloud, Azure, and air-gapped installations. The vendor advertises 138 models for embeddings, reranking, OCR, document-to-Markdown conversion, structured extraction, guardrails, and agent loops. A stateless gateway, shared queue, worker pools, batching, model stacking, and autoscaling share GPUs across workloads. Listed runtimes include vLLM, SGLang, Text Embeddings Inference, llm-d, and NVIDIA Dynamo. An OpenAI-compatible endpoint eases integration.

Pricing and commercial terms: Self-hosted SIE is described as always free, but customers pay for Kubernetes, GPUs, storage, networking, monitoring, upgrades, and on-call operations. Managed hosting and an agent plug-in are marked upcoming. The pricing route returned 404, so no managed-service amount is asserted.

Competitive context: SIE consolidates several small-model and document workloads on one private cluster. Chroma, LanceDB, and Weaviate primarily anchor retrieval storage, while LangChain orchestrates applications; SIE supplies the inference layer they can call. Relevant alternatives include Weaviate, Chroma, LanceDB, LangChain. Choose based on deployment control, integration effort, measurable task quality, and total operating cost rather than a polished demonstration.

Practical strengths include Data can remain in customer infrastructure; One cluster supports mixed model workloads; Apache 2.0 avoids license fees; Air-gapped deployment is supported. Important limitations are Customer owns Kubernetes and GPU operations; Managed products were marked upcoming; No managed pricing was published; Vendor benchmarks require local reproduction. These tradeoffs matter because AI output can look plausible while still being incomplete. Keep human approval around publishing, financial entries, purchases, account changes, deletion, or other consequential actions. Confirm retention, deletion, exports, role controls, logs, model-training terms, support commitments, and regional availability before production.

A useful pilot should cover at least 20 representative tasks and include normal cases, ambiguous inputs, stale data, permission failures, retries, and cancellation. Record successful completion, factual accuracy, p95 latency, human correction time, interventions, and end-to-end cost. For retrieval products, measure recall and citation accuracy; for browser agents, test dynamic pages and expired sessions; for accounting or market content, reconcile outputs against primary records. Assign a named owner, preserve source evidence, and define rollback before enabling write actions. The product belongs on a shortlist only when the pilot shows repeatable value under realistic failure conditions. Before rollout, document the current baseline, expected savings, acceptable error rate, escalation owner, and stop conditions. Recheck vendor pricing and product limits at purchase because plans, quotas, integrations, and model behavior can change after this research date.