saved
Ultramarine 1 Flash: an open model for knowledge work
Hraness wrote this summary from a saved copy of the source. Quotations are taken word for word from the source.
gist
Daanish Khazi announces Paper Instruments' Ultramarine 1 Flash, an open-source model post-trained for finance, consulting, operations, and law work that he says nears Claude Fable 5 at a fraction of the cost. It was post-trained by five people for about $300k, with training code and a tech report released and weights promised within two weeks.
ideas
- Frontier quality, small budget. A team of five post-trained the model for roughly $300k all-in; the launch chart puts it at $0.13 per task against $1.26 to $8.19 for closed peers.
- Cost claims vary by framing. The thread says 60x cheaper than Fable 5, while the video's output-price chart shows $0.50 against $50.00 per million tokens and labels it 100x.
- Taste through non-verifiable RL. Prolonged RL with rubric judges is credited with teaching professional standards such as visual hierarchy and plain prose.
- Open training stack. Released code includes the harness, rollout infrastructure, rubric judge, and Paper Office packages for Excel, PowerPoint, and Word files.
- Positioned as a drop-in. It targets AI companies serving banks and law firms that now run Claude or GPT in main loops or sub-agents.
quotes
“It was post-trained by a team of 5 for ~$300k all-in.”
“Through prolonged training, the model learnt professional standards: visual hierarchy, straightforward prose, etc.”
“UM Flash is an obvious drop-in replacement to both improve quality and pass on cost savings to your customers.”