Heabsy
by FEYA, s.r.o.
OpenAI- and Anthropic-compatible inference API for open-weight models, with an EEA-only tier
Heabsy is an OpenAI- and Anthropic-compatible inference API for open-weight models, run by FEYA, s.r.o., a small Slovak company. It is for European developers and companies that want one EU counterparty, one DPA and per-token billing, with a small tier of models processed only inside the EEA.
Heabsy is the brand of FEYA, s.r.o., a Bratislava company registered in September 2014 whose managing director is Kostiantyn Fomichov. It runs an AI implementation business for mid-sized firms in Austria, Germany and Slovakia and, alongside it, an inference API for open-weight models at api.heabsy.com/v1. The API accepts OpenAI SDK calls and has an Anthropic-compatible endpoint, so Claude Code, Cursor and similar clients can point at it by changing the base URL; streaming, tool calling and structured output are supported. The catalog has two tiers. The EEA tier, a handful of models led by Qwen3.8 27B served in NVFP4, runs on GPUs inside the European Economic Area, and Heabsy says prompts and completions are processed in memory and never logged. The site describes these as its own machines, while its privacy policy says the capacity is rented from third-party GPU providers, including GPU-marketplace hosts, located in the EEA. The routed tier, more than 25 models from DeepSeek, Moonshot, Z.ai, OpenAI, Meta, Google, Qwen and NVIDIA, is fulfilled through OpenRouter and may be processed outside the EEA; the catalog labels each model's processing region. Pricing is per million tokens with no minimum spend, from $0.04 input on GLM-5.3-Flash to $0.40 input, $0.05 cached and $3.00 output on the EEA Qwen model (price list of 12 September 2026). The models are also resold through gateways such as NanoGPT and LLMTR. Slovakia's company register lists FEYA with one employee.
A European developer or small engineering team that needs open-weight model inference billed by an EU company under one DPA, and can live with a single-vendor dependency on a very small provider.
One OpenAI- and Anthropic-compatible key for 30-plus open models on one EU invoice, with a few models processed only inside the EEA and the region labelled per model.
At a Glance
- Category
- AI Models & APIs
- Pricing
- Usage-based
- Target Market
- Enterprise Developers, CTOs, Startups, AI engineering teams
- Deployment
- API-based, Cloud-only
- Founded
- 2014
- Headquarters
- Bratislava, Slovakia
- Team Size
- 1-10
Key Features
- ✓OpenAI-compatible API
Chat completions, SSE streaming with usage, tool calling and structured output work with existing OpenAI SDKs by changing the base URL.
- ✓Anthropic-compatible endpoint
Lets Claude Code, Cursor and other Anthropic-protocol clients run against open-weight models hosted by Heabsy instead of a US provider.
- ✓EEA processing tier
A few models, led by Qwen3.8 27B, run on GPU capacity inside the EEA, and Heabsy says prompt and completion text is never written to logs.
- ✓Routed open-model catalog
More than 25 open models bought from global providers through OpenRouter, billed by Heabsy on one EU invoice, with each model's processing region labelled.
- ✓Prefix caching
Routes a conversation to the machine holding its prefix and bills cached input at a lower rate, which cuts cost on agent workloads that reread context.
- ✓Per-key budgets and limits
The console issues multiple API keys, each with its own spending budget and RPM or TPM rate limits, and reports fresh, cached and output tokens per request.
Capabilities
Use Cases
- •EU-hosted coding agent backend
A developer points Claude Code at the Anthropic-compatible endpoint and runs Qwen3.8 27B in the EEA tier, paying mostly cached-input rates on long sessions.
- •Single EU contract for open models
A European company replaces several US inference contracts with one DPA and one invoice from a Slovak company covering DeepSeek, Llama, Gemma and Qwen.
- •Low-cost extraction and routing
A team sends classification and field-extraction calls to small routed models such as gpt-oss-20b or Llama 3.1 8B at a few cents per million input tokens.
- •Model comparison during prototyping
Engineers switch between open models by changing one model string while keeping the same SDK, key and billing, then pick a model for production.
Ideal For
Best For
- ✓EU teams that want an EU-incorporated counterparty and a GDPR Article 28 DPA for open-model inference
- ✓Developers running Claude Code, Cursor or Cline against open-weight models through an Anthropic-compatible endpoint
- ✓Agentic coding workloads that reread long context, where cached input is billed at $0.05 per million tokens on the EEA Qwen model
- ✓Prototyping across many open models (DeepSeek, Kimi, GLM, Llama, Gemma, Qwen) behind one key with per-token billing
- ✓Security research and red-teaming that needs low-refusal model variants, which Heabsy lists separately
Not Ideal For
- ✗Enterprises that need SOC 2 or ISO 27001 certification, which Heabsy does not hold (its homepage lists SOC 2 as a roadmap item)
- ✗Buyers who require all processing inside the EEA across a broad model choice, because only a few models run in the EEA tier and the routed tier goes through OpenRouter
- ✗Procurement teams that need a vendor with financial depth and support staff; the Slovak register lists FEYA with one employee
- ✗Teams that also need embeddings or raw GPU rental, which ComputePrices lists as unavailable
Deployment
Market Analysis
Pros
- ✓Drop-in OpenAI and Anthropic compatibility, confirmed by third-party gateways (NanoGPT, LLMTR) that resell its models
- ✓Transparent per-token price list with cached-input pricing and no minimum spend
- ✓Per-model processing labels, plus a privacy policy that names its sub-processors, including OpenRouter
- ✓NanoGPT independently observed about 176 tokens per second and 0.9 s time to first token on Qwen3.8 27B, matching Heabsy's own throughput claim
Cons
- ✗The site says the EEA tier runs on Heabsy's own GPUs, but its privacy policy says that capacity is rented from third-party GPU providers, including marketplace hosts
- ✗Only a few models are processed strictly in the EEA; most of the catalog goes through OpenRouter, a US-based router, and may run outside the EEA
- ✗No SOC 2 or ISO 27001 certification, and no embeddings endpoint (per ComputePrices)
- ✗Very small operator: the Slovak register lists one employee, and the inference API sits beside a consulting business
- ✗No user reviews on G2, Capterra, Product Hunt, Hacker News or Reddit, so production reliability is unproven by third parties
Pricing
Routed models (pay per token)
From $0.04/M input tokens
- ✓More than 25 open models
- ✓May be processed outside the EEA
- ✓No minimum spend, no subscription
EEA tier: Qwen3.8 27B
$0.40/M input, $0.05/M cached, $3.00/M output
- ✓EEA-only processing
- ✓Zero data retention (vendor-stated)
Dedicated EEA deployment
Contact for pricing
- ✓Open model brought onto dedicated EEA capacity for a steady workload
Heabsy publishes per-million-token prices for input, cached input and output on every model at heabsy.com/models, with no minimum spend or subscription (price list dated 12 September 2026). Routed models start at $0.04 per million input tokens and reach $2.20 input and $11.00 output for Kimi K3; the EEA Qwen3.8 27B costs $0.40 input, $0.05 cached and $3.00 output. Accounts are opened on request with a spending limit, ComputePrices lists no free tier, and dedicated EEA deployments are quoted on request. Resellers such as NanoGPT add their own markup.
Security & Compliance
Sources
This page was written from 7 sources, 4 on domains other than heabsy.com.
Stay Ahead of the Curve
Weekly enterprise AI insights for technology leaders. No spam, no vendor pitches—unsubscribe anytime.
SubscribeRelated Products
Liquid AI LFM3
Open-weight 2B text, vision, audio and decision models that run on phones, laptops, cars and private servers
Recursive Flint
Apache 2.0 open-weight coding model (40B MoE, 2.5B active) built to run coding agents locally on a laptop
Arena AI Evaluations
Paid model evaluations built on the Arena leaderboard's crowdsourced human votes, for enterprises choosing an LLM
Kimi K3
2.8T-parameter open-weight MoE model with a 1M-token context, priced at $3/$15 per million tokens