Heabsy AI Inference

Heabsy AI Inference (api.heabsy.com, console at platform.heabsy.com) is the API arm of FEYA, s.r.o., a Slovak company (Bratislava registered office, IČO 47887541) that otherwise sells bespoke RAG/agent implementation projects to mid-sized regulated companies in Austria and Germany. The API is OpenAI- and Anthropic-compatible — point an existing SDK at api.heabsy.com/v1 with a model ID from the catalog and it works. One model, Qwen3.8 27B, is the flagship: Heabsy operates the GPUs for it itself, inside the EEA (datacenters in NL, DE, DK, NO, BG, per the live catalog API). Everything else — about 30 other open-weight models from DeepSeek, Qwen, Mistral, Meta, Google, NVIDIA and others — is routed through third-party global providers under Heabsy's own contract: one DPA, one invoice, but not Heabsy's own hardware.

The API is billed per million tokens, with a separate price for input, cached input and output on every model — confirmed against the live price list at heabsy.com/models (updated 12 September 2026) and the machine-readable catalog at api.heabsy.com/openrouter/v1/models. Prices range from $0.04/M input on the cheapest routed model to $2.20/M input on Kimi K3; the flagship Qwen3.8 27B is $0.40/M input, $0.05/M cached input, $3.00/M output. No minimum spend, no subscription. The €5K (Starter) and €25K (Growth) figures that also appear on heabsy.com are a separate product — fixed-price AI implementation consulting projects (document audits, RAG deployments, integration work) — and have nothing to do with the inference API's token pricing.

Heabsy's own security page states plainly, under a “Not certified” label: “Are you ISO 27001 or SOC 2 certified? No.” It adds that it can help a customer build toward their own certification but does not claim either standard for itself. There is no 30-day or other free trial advertised anywhere on the platform, models or pricing pages — access is a self-service signup and pay-per-token from the first request.

Heabsy's own FAQ page states it directly: “Does not claim that every model reachable through its API is processed in the EU: routed catalogue models may run outside the EEA, and each model page says so.” The live model table bears this out — most routed models (DeepSeek, Llama, GLM, Nemotron, Kimi K3 and others) are tagged “Outside EEA”; a handful of routed models, including some uncensored/fine-tuned variants, are tagged “EEA only.” A buyer who needs strict EEA-only processing across every model they call should check the Processing column on heabsy.com/models before relying on any single model, rather than assuming the EEA promise extends to the whole catalog because the flagship is EEA-hosted.

The public models page lists 31 models. Querying the same live catalog endpoint the site uses (api.heabsy.com/openrouter/v1/models) returns several additional, currently-queryable model IDs not listed on that page: refusal-removed (“uncensored,” “obliterated,” “abliterated”) variants of Qwen3.8 27B, DeepSeek-V4.1-Flash, GLM-5.3-Flash and MiMo V2.6 Flash; an in-house “Heabsy V1” assistant model; and a model called “CyberHeabsy,” whose own API description markets it for “red-teaming, security auditing, reverse engineering, and advanced malware analysis.” None of these appear on heabsy.com/models or elsewhere on the public site. This isn't a security flaw — the key only reaches what it's issued for — but a buyer reviewing the published catalog before signing is reviewing a narrower list than what the API key can actually call.