AI News
All stories
Reported

Ramp Labs says its post-trained Qwen 3.5 model beat Opus 4.6 on finance questions

Ramp Labs' Alexander Shevchenko says a post-trained Qwen 3.5 model beat Opus 4.6 on financial question answering at Haiku-level latency.

AI in Practice · · 1 min read

Alexander Shevchenko of Ramp Labs says a specialist Qwen 3.5 model built for Ramp Sheets beat Opus 4.6 on financial question answering while running at Haiku-level latency. In a conference talk, he argued that application companies should keep doing applied research instead of waiting for the next frontier model. He described the Qwen result as one of three projects Ramp says are now in production. The description of the talk does not say which financial questions were used, how answers were scored, or how large the Qwen model is, so the comparison with Opus 4.6 rests on Shevchenko's account alone.

Shevchenko also described KV cache cartridges, which he says pack 20 to 30 times more context into an agent ahead of time for tasks such as month-end close. The third project put swarms of 10,000 parallel agents to work pen-testing Ramp, which he says turned up seven high-severity vulnerabilities. He also said cheaper open-weight models caught nearly as many vulnerabilities in those swarms.

What changed

A post-trained Qwen 3.5 model built for Ramp Sheets reportedly beat Opus 4.6 on financial question answering at Haiku-level latency, according to the speaker's own account.

Type: Research · Applies to: Qwen 3.5, Opus 4.6, Ramp Sheets

Why it matters

Companies building agents may get better latency and cost on narrow tasks from a smaller post-trained model than from a larger frontier model. These results are the speaker's own claims from a talk and have not been independently verified.

What you can do

Test a smaller post-trained model against a frontier model on your own finance or back-office tasks before assuming the largest model is required.

Original source

We have no link to the original source for this story yet.

Where we saw it

1 source

AI Engineer

AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it. How it works