Ramp says its 35B model beat Opus on finance QA

Before the November launch of Ramp Sheets, Ramp's finance agents leaned on frontier models alone. Alexander Shevchenko told AI Engineer that trade-off has flipped. Ramp Labs demoed 10,000 parallel agents that pen-tested Ramp's own platform and found 7 high-severity vulnerabilities reachable by remote email.

In 2023 the stack was GPT-4 with strict guardrails, no backtracking and brittle prompts, Shevchenko said. By November last year Ramp Sheets shipped as an early spreadsheet agent ahead of frontier labs, and the team then hit friction when Opus 4.6 planned and retrieved across large sheet packs. Labs paired with Prime Intellect to post-train a 35 billion parameter Qwen 3.5 retrieval model, Shevchenko said, and the smaller system beat Opus 4.6 on exact-task accuracy for financial question-answering while running at far lower latency.

The second bet is KV cache cartridges. The technique compresses 20 to 30 times more context into the cache offline, Shevchenko said, so month-end close books can be prepared overnight. The agent opens on the first of the month with ledgers, aging reports and documents already loaded, no agentic retrieval needed.

Shevchenko said the team could not wait for the next frontier jump: "We need to serve our financial professionals now." That urgency shaped Ramp Inspect, the cloud harness that mapped work across the swarm, deduplicated tickets with LLMs and adversarially reproduced each finding to filter noise. One chain required stitching multiple safe components to achieve Lua script injection via email, Shevchenko said, and open-weight models at five times lower cost delivered near-frontier detection when run as a swarm.

The demo leaves gaps a practitioner would notice. Shevchenko named no CVE identifiers, no affected version numbers and no patch timeline beyond saying the seven issues were patched, and the attack required only a remote email with no local access. The same economics still apply: token cost is the only filter on false positives, which favors cheap parallel runs over single expensive traces.

Ramp said at the time of its $1.85B round it had passed $2 billion in annualized revenue with positive cash flow, per AI Funding Tracker, and Shevchenko said more routing strategies for Ramp Router are scheduled to launch later alongside additional Ramp Labs posts on Portal and self-improving monitors.