Skip to content

status.openkey.ai is live — independent, worldwide uptime proof See it now ->

openkey

Sao10K

4 models built

Sao10K is an independent fine-tuner, not a corporate lab — no funding rounds or research papers, just a track record of Llama-based fine-tunes built for roleplay and character-driven chat. The lineup runs four models across three series: Euryale, Lunaris, and Hanami. All four are fine-tunes of Meta's Llama 3 family (3, 3.1, and 3.3), which means the strategy isn't building new architectures — it's tuning existing open weights for a specific use case Meta's instruct models don't target well. On OpenKey you get all four through the same API key as every other model, billed at provider price plus a flat 3% fee.

Lineup

Start with Llama 3.3 Euryale 70B ($0.65/$0.75 per M input/output) — newest base model, 131K context, and the cheapest of the 70B tunes. Llama 3.1 Euryale 70B v2.2 ($0.85/$0.85) is the prior version, same 131K context, worth using only if you've pinned to it for consistency. Llama 3.1 70B Hanami x1 ($3.00/$3.00) is the priciest by far and has a much smaller 16K context — pick it only if Hanami's specific tuning is what you need. For cheap, high-volume roleplay or testing, Llama 3 8B Lunaris ($0.04/$0.05) is the obvious default; it's a fraction of the 70B cost with an 8K window.

Prices per 1M tokens, flat 3% fee included.

Questions

What does Sao10K pricing look like on OpenKey?
Four models, all paid — no free tier here. Prices span $0.04–$3.00 per M input tokens depending on the model: Lunaris 8B is $0.04/$0.05 (in/out), the Euryale 70B models run $0.65–$0.85, and Hanami x1 tops out at $3.00/$3.00. OpenKey charges provider list price plus a flat 3% fee, so Lunaris 8B input works out to $0.04 x 1.03 ≈ $0.0412 per M tokens.
Which Sao10K model should I use by default?
Llama 3.3 Euryale 70B. It's the newest of the four (Dec 2024), has the largest context window at 131,072 tokens, and at $0.65/$0.75 per M tokens is cheaper than both the older Euryale v2.2 and the pricier Hanami x1. Only drop to Lunaris 8B if you need lower cost and can accept an 8K context limit.
Is there API access to Sao10K models beyond Hugging Face?
Sao10K publishes weights on Hugging Face; OpenKey exposes all four fine-tunes — Lunaris 8B, Euryale 70B v2.2, Euryale 70B (3.3), and Hanami x1 — through one API key alongside the other 325+ models on the platform, so you don't need separate hosting or provider accounts to call them.