Heabsy

Heabsy sells access to AI models rather than a model of its own: an OpenAI- and Anthropic-compatible API, built so a team already using the OpenAI SDK or running Claude Code, Cursor or Cline can point their client at a new base URL and keep everything else unchanged. The company, FEYA, s.r.o., is registered in Bratislava, Slovakia, with no US parent, which keeps the operator itself outside the reach of the US CLOUD Act.

The distinction that matters most on this page is between the flagship model and the rest of the catalog.

Heabsy's own model, Qwen3.8 27B, runs on dedicated GPUs in EEA data centres, and its Data Processing Agreement is specific about what that buys: prompts and completions are never written to logs, nothing is used to train any model, and no request content leaves the EEA.

Everything else in the catalog — models from Z.AI, OpenAI's open weights, Mistral, DeepSeek, Google and Meta — is fulfilled through global providers under Heabsy's own European contract, on one invoice and one DPA, but the compute for those may run outside the EEA under that provider's own policy.

On performance, Heabsy publishes numbers rather than asking to be taken on faith: roughly 176 tokens per second on the flagship model, about double the fastest public provider of the same model on the OpenRouter showcase, with a time to first token of 0.3–0.9 seconds.

The gain comes from a trained speculative-decoding draft running on dedicated hardware, not from quietly serving a smaller model.

A prefix cache keeps a conversation pinned to the machine that already holds its context, which the company says covers 91–97% of tokens on agentic workloads and cuts a repeated 34,000-token prompt from 42 seconds to 0.48 — a difference an agent that iterates ten times on the same context will feel directly.

What Heabsy does well

• Slovak company with no US parent, and the flagship model runs on dedicated GPUs in EEA data centres with a documented zero-retention policy
• OpenAI- and Anthropic-compatible: point an existing SDK or agent tool at a new base URL, nothing else changes
• Published, verifiable performance numbers (throughput, time to first token, cache hit rate) rather than marketing claims
• Pay-as-you-go with no minimum spend and no subscription, and per-key budgets with early rate-limit signals instead of a silent queue

Where Heabsy falls short

• Only one model (Qwen3.8 27B) carries the full EEA-only, zero-retention guarantee; the rest of the catalog depends on third-party providers whose compute may sit outside the EEA
• Young platform with no published founding date or team page
• No free tier, and the flagship model's 262,144-token context is smaller than some routed alternatives in the catalog

Standout feature. The prefix cache: on agentic coding workloads where most tokens are rereads of context the model has already seen, Heabsy reports 91–97% of tokens served from cache rather than recomputed, and bills cached input at a fraction of the fresh-token rate.