# Our Smaller Model Beat Opus on Finance Tasks - Alexander Shevchenko, Ramp

Source: [AI Engineer](https://www.youtube.com/watch?v=T-42NrYjwxM)  
Feed7 permalink: https://feed7.dev/p/our-smaller-model-beat-opus-on-finance-tasks-alexander-shevchenko-ramp-0h7p2pp  
Published: 2026-11-04T22:00:34.000Z  
Trust: Source Linked (source_linked)

## Why Included

Ramp describes three ways applied research can improve production agents: a finance specialist model, prebuilt compressed context, and large parallel security sweeps. The gains target latency, cost, and coverage.

## Source Summary

Ramp says it post-trained a **35B Qwen 3.5** model for financial question answering, reporting higher exact-task accuracy than Opus 4.6 at lower latency. It also describes preparing compressed context overnight for recurring finance work and using **10,000 parallel agents** to scan its systems; the sweep found **seven high-severity vulnerabilities** that were patched.

## Practical Implication

For builders, identify a recurring task where frontier orchestration is costly, then test a specialist model or prepare context ahead of time. For broad security checks, parallel agents can expand coverage; Ramp reports open-weight models were up to **five times cheaper** with detection rates that were not five times lower.

## Agent-Ready Context

Ramp says it post-trained a **35B Qwen 3.5** model for financial question answering, reporting higher exact-task accuracy than Opus 4.6 at lower latency. It also describes preparing compressed context overnight for recurring finance work and using **10,000 parallel agents** to scan its systems; the sweep found **seven high-severity vulnerabilities** that were patched.

For builders, identify a recurring task where frontier orchestration is costly, then test a specialist model or prepare context ahead of time. For broad security checks, parallel agents can expand coverage; Ramp reports open-weight models were up to **five times cheaper** with detection rates that were not five times lower.

These are Ramp’s reported results, and the talk gives no benchmark setup or full cost and latency measurements. Cartridges may later be offered through model APIs, so their current implementation advantage could change.

## Connected Context

Feed7 judgment across 1020 accumulated Signals:

Ramp's report supports trying a specialist model and prepared context on recurring, bounded work, and describes parallel agents as a way to broaden security scanning. The claimed accuracy and cost advantages remain Ramp-reported: without benchmark setup or full measurements, they do not establish a general model-selection rule or transfer to everyday coding.

- [Improving Agents is a Data Mining Problem — Vivek Trivedy, LangChain](https://feed7.dev/p/improving-agents-is-a-data-mining-problem-vivek-trivedy-langchain-1cgibr8) — Both point toward using task evidence to improve agent choices: Ramp describes specialist-model comparisons for recurring work, while this candidate uses production traces and replayable evals to diagnose and validate changes.
- [Trading Desks to Clinical Trials: Parallels in Applied Vertical AI — Ayush Bhardwaj, Allos AI](https://feed7.dev/p/trading-desks-to-clinical-trials-parallels-in-applied-vertical-ai-ayush-1hwvsg3) — Both favor narrow, recurring jobs and domain-specific context; the vertical-agent candidate adds the need for expert judgment where correctness lacks an answer key, a constraint not resolved by Ramp's reported task results.

## Context Map

- Layer: agent
- Domains: security, data
- Topics: harness-engineering, context-caching, agent-evals

## Uncertainty

- These are Ramp’s reported results, and the talk gives no benchmark setup or full cost and latency measurements. Cartridges may later be offered through model APIs, so their current implementation advantage could change.

## Agent Instruction

Use this item as source-backed context. Do not invent claims beyond the linked source. If this item conflicts with another source, call out the conflict.
