Fireworks AI Co-Founder Says Open Models Trail Frontier by Months, Addresses Two Breaches
Source: fireworks.ai

Image: kewl on Pixabay
Fireworks AI co-founder Chenyu Zhao said on November 19, the day Anthropic went public, that most companies can use open-weight models instead of the priciest frontier ones. He also addressed two security incidents that the inference provider describes in its security bulletins. Anthropic's IPO priced at $83 a share, raising about $81.3 billion.
Zhao puts the open-model gap at three to six months
In an interview, Zhao said open-weight models trail frontier models by at most six months, probably closer to three. He said companies spending $10,000 to $100,000 a month on tokens should look at open models. He described a "two-model world" in which frontier and open models run side by side. Fireworks says it serves open models at five to 10 times lower cost than closed alternatives.
That cost figure is the company's own. No workload, model pair or pricing basis was given with it. The saving for any team depends on the frontier model it uses today, for example Gemini 4 Neon at $3 per million input tokens. The three-to-six-month figure is also Zhao's estimate, and the real gap varies by model and by task.
Fireworks sells hosted inference, not local setups
Fireworks announced a $1.505 billion Series D at a $17.5 billion valuation in July. It said it had passed $1 billion in annualized revenue run rate. The company pitches customers on adapting open models to their own data and running them next to closed ones. Its customers include Uber and Shopify, according to that announcement. Zhao's argument is therefore about open models served from Fireworks' cloud.
Bulletins describe two separate intrusions
A bulletin dated November 17 says the activity covered in earlier updates came from two separate incidents. Fireworks said the first involved unauthorized use of internal credentials and access to environment variables in evaluation jobs for a limited number of accounts.
In the second, Fireworks says an unauthorized party reached its development machines at a third-party cloud provider that had suffered its own incident. The party then used that access to reach part of its production systems. Fireworks makes these claims about the impact:
- Account and billing information was accessed for accounts created before October 30.
- Datasets, models or secrets were accessed for a small number of customers.
- Payment card numbers were not affected, because Fireworks does not store them.
- Inference, Training SDK and Nexus showed no impact. Fireworks says all three are zero data retention by default.
Fireworks also says it revoked and replaced credentials and service accounts, removed unneeded elevated privileges, and notified affected customers directly. The bulletins say a forensic investigation is under way. Zhao addressed the incidents, which have cost Fireworks some customers' trust.
No model, licence or hardware figures were given
Zhao named no specific model, so there is no parameter count, licence, context length or memory requirement to report. Teams that want to keep prompts and data on their own hardware can run open-weight models locally with Ollama, LM Studio, llama.cpp, or MLX on Apple silicon. That removes third-party hosting from the picture. The size of the model then decides how much memory or GPU capacity the machine needs, and Zhao's remarks do not say which models he means.
Open questions for buyers of open-model hosting
Several things are still unclear. Fireworks has not said which models beat which frontier ones on which tasks. The five-to-10-times saving has no published methodology behind it. The bulletins describe the number of customers whose datasets, models or secrets were accessed only as small. Customers weighing a switch from a frontier API will need to compare Fireworks' prices against their current bill. They should also read the bulletins to see whether their own account is among those flagged. Fireworks says accounts that did not receive a notice were not identified as affected so far.
Featured image: kewl on Pixabay
Get the week's tech news by email
One email a week with Trendblog's stories on AI, hardware, apps and crypto.
We'll email you to confirm your address. Unsubscribe any time. Privacy policy