We merge models
Failure Model Labs is a small lab in Singapore. We build models out of other models — at inference across closed frontier APIs, and in the weights of open ones.
No single model is best at everything
Every frontier model has a shape: problems it solves cleanly and problems it fumbles. Those shapes do not line up. Where one model is confidently wrong, another is often right, and a careful check can tell which.
That gap is where we work. Tempest 1 puts four frontier models on the same problem under different lenses and settles their disagreements by checking. In the lab, we pursue the same idea in open weights.
The work
API fusion
Failure Tempest 1. Route, brief, lensed drafts, fusion, audit.
Expert transplant
A frozen spine with expert layers transplanted from other mixture-of-experts families through learned adapters.
Weight merging
Several fine-tunes of one base model merged into a single checkpoint.
Rules we keep
Claim nothing we did not measure
No borrowed benchmarks and no rounded-up results. Until our own runs exist, the score is a dash.
Publish the losses
When we evaluate, the prompts where fusion made things worse go out next to the ones where it helped.
Bill for what you can see
Every brief, draft and audit streams to you as reasoning. You pay for tokens you send and tokens you receive, thoughts included — nothing hidden.
Do not train on your traffic
Prompts sent to the API are not training data. Request content lives only in caches measured in hours.
A small team
Failure Model Labs was founded by Zandy. We are small on purpose: few enough people that the person who answers your email is the person who wrote the router.
Where we are aiming
Our target for Tempest 1 is #1 on Artificial Analysis. It is a target, not a result. We will say when we have run it, and we will show what we found either way.