typescript · llm · telegram
Migrating a production bot to Nebius Token Factory
The one-line diff that moved my live Telegram translation bot to Token Factory, plus a measured three-model comparison on the real workload.
The setup
I run a translation bot in live multilingual Telegram groups (English,
Italian, Portuguese). Token Factory is OpenAI-compatible, so the
migration was a baseURL and API-key change — the handlers, prompts,
streaming, and token accounting kept working on the official OpenAI
Node SDK.
The comparison
I ran the bot’s core translate call over ten real-shaped group messages across three price tiers and published the numbers: median time to first token, total latency, tokens out, and cost per thousand messages. The surprise: the flagship model was also the fastest, because it is the only one that does not burn hundreds of reasoning tokens before answering — which collapses the price gap on paper.
Why it reads like DevRel
Latency and tokens are measured; quality verdicts are labeled as my own human eval (native Italian speaker, fluent Portuguese). The repo shows the actual diff and honest notes on the experience, not a vendor pitch.