How an American AI Company Uses Chinese Open-Weight Models
Source
omny.fm
Date
Why it matters
Practitioners choosing between closed frontier APIs and open-weight models get a first-hand account of an AI company's use of Chinese open models and what it implies for pricing power and model commoditization.
Key takeaways · AI-distilled
Minerva, a marketing-automation startup, builds its product with frontier closed models such as Claude, then serves repeatable tasks on cheaper models, judging them by cost per completed task rather than cost per million tokenThe chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.Full definition →, its co-founder told Odd Lots.
Minerva fine-tuned Qwen on a few thousand graded examples of complex audience requests. He says the tuned model is faster than the OpenAI model it replaced for turning those requests into SQL and better at the queries customers actually send.
He ranks the reasons to use open weightsA model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.Full definition → as cost first, then control over data, so a model improves on a company's own use, then leverage, since an in-house model pushes labs toward value-based pricing.
fine-tuningTaking a trained model and training it a bit more on your own examples so it gets better at one specific job.Full definition → is now a commodity service, he says: Fireworks charges about $7 to $9 per million tokens, and some good fine-tunes cost about $90 for the final run. The tuning does not carry over to a new base model version, so retraining is an ongoing cost.
His main worry is refusal-stripped open models, which he called "obliterated" models. He says an uncensored DeepSeek he tried costs about 30 cents per million blended tokens, about 1/100th of Fable, and will follow any instruction.
Terms in this piece · Glossary
open weights — A model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.
token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
fine-tuning — Taking a trained model and training it a bit more on your own examples so it gets better at one specific job.