The use case: how Paper Instruments deploys Fireworks AI
Paper Instruments trains its Ultramarine 1 Flash model, an open model for professional analysis, writing, and presentations, using reinforcement learning on the Fireworks Training API and Fireworks inference. Fireworks provides rollout generation, GPU tuning, and training/inference numerical alignment, and serves the resulting model through serverless inference.
More reported results
Figures published in Fireworks AI’s case study. They are not independently measured by usedby.
Ultramarine improved on GLM-5.3 Flash across all six reported knowledge-work benchmarks, gaining 4.1 points on average.
On Paper Instruments’ internal PI Knowledge Work benchmark, Ultramarine’s score rose from 52.0 to 59.3, a 7.3-point improvement.
Total development cost was approximately $300,000, including the data creation and experimentation that preceded that run.
After developing its curriculum through smaller-model experiments, the team completed the GLM-5.3 Flash training job on Fireworks in approximately two weeks.
In the customer’s words
We retained full control over sampling and shaping the reward. We could focus on the RL science without having to think about kernel-level optimization, sizing, and parallelism.Other AI tools used by Paper Instruments
Other public relationships recorded on usedby.
Explore Paper Instruments’s AI stack
Other companies using Fireworks AI
Customers across industries, with public evidence.
See all Fireworks AI customers
Go straight to the source
Based on Fireworks AI’s published case study. We check the public evidence; the results are reported by the vendor.
fireworks.ai- Source published
- Oct 28, 2026
- Last checked
- Oct 31, 2026
Open source capture What the story claims, and what we checked
What we compared with the page.
- The figure: ~98%Checked
Printed word for word on the page, near the name of Paper Instruments.
- The passage quoted aboveChecked
Copied word for word from the page, near the name of Paper Instruments.
- The numbers in our summaryChecked
Each one is printed on the page.
- The publication dateChecked
Read from the page’s own metadata, never guessed.
- Paper Instruments uses Fireworks AIChecked
Confirmed line.
- The result itselfNot checked
We quote it; we did not measure it.
On Paper Instruments’ internal PI Knowledge Work benchmark, Ultramarine 1 Flash scored 59.3 versus Claude Fable 5’s 57.8, at $0.13 versus $8.19 per task.