Open weights,
served straight.
Kiminohappydays runs the leading open-weight language and audio models on dedicated hardware in the EU and US. One OpenAI-compatible endpoint, billed per token, with nothing trained on your traffic.
No card to start · Frankfurt + Ashburn regions · SOC 2 in progress
Why Kiminohappydays
The open ecosystem, without the ops.
Open-weight models caught up. Running them well — batching, quantisation, tail latency, region residency — is still a job. We do that part, and keep the interface identical to what you already use.
Compatible
Drop-in schema
Chat, embeddings, reranking and audio all follow the OpenAI request shape. Change the base URL and key; keep your SDK.
Private
Nothing retained
Prompts and outputs are never logged beyond the request or used for training. EU traffic stays in the EU when you pin the region.
Honest cost
Per token, hard caps
No seats, no minimums. Spend caps fail closed with a 429 instead of quietly running up a bill.
Selected models
Frontier open weights, plus audio.
A curated set rather than a dump — the models we can serve at low tail latency and stand behind in production. Full catalogue with specs and licences on the models page.
Qwen3-235B-A22B
Apache 2.0Alibaba · mixture-of-experts
Flagship general and reasoning model. 235B total parameters, 22B active per token, strong multilingual and tool use.
DeepSeek-R1
MITDeepSeek · reasoning
Explicit chain-of-thought reasoning for maths, logic and hard code. Emits thinking tokens you can show or hide.
Whisper large-v3
MITOpenAI · speech-to-text
Robust transcription and translation across 99 languages, with word-level timestamps. Billed by the minute of audio.
First request
Point your client at us.
If your code already speaks the OpenAI schema, this is the whole migration.