Paper Instruments Opens Ultramarine 1 Flash, Claims 98% Lower Cost

Can an open model do the work of an analyst, a lawyer or a consultant, and go head to head with closed flagships on both quality and price? That is the story Paper Instruments is telling. The company has just released the open weights of Ultramarine 1 Flash and put the model on Fireworks and Baseten.

An open model built for knowledge work

Ultramarine 1 Flash is built on Z.ai's GLM-5.3 Flash with reinforcement-learning post-training. It has 320B total parameters and 18B active, and takes text, image and video input. Baseten lists the license as a modified MIT license.

The training idea is to teach the model what good professional work looks like. The team built synthetic tasks across 19 practice areas and 88 environment types, with task-specific rubrics that reward analytical judgment, clear writing and appropriate presentation. The company says that, paired with the right agent harness and a code-execution environment, it can produce Excel, PowerPoint and Word files.

Two ways to run it

Fireworks offers pay-per-token serverless inference. Its page lists $0.15 per million input tokens, $0.03 for cached input and $0.50 for output, with a context window of about 1M tokens. Baseten offers dedicated deployments, which suit teams that need tighter latency and isolation.

Putting the 98% in context

The headline numbers are eye-catching. On Paper Instruments' own PI Knowledge Work benchmark, Ultramarine scored 59.3 against 57.8 for Claude Fable 5, at $0.13 versus $8.19 per task, about 98% lower. The company also says it gained 4.1 points on average over its base model across six knowledge-work benchmarks.

But the benchmark is the company's own, and Fireworks' case study calls it an internal knowledge-work benchmark. So 59.3 versus 57.8 shows the model does well on tasks this company chose, which isn't the same as winning on your workload.

Why it's worth watching

Open models have caught up quickly in coding and reasoning, but professional-services work, where taste matters, still mostly belongs to closed models. Paper Instruments takes an existing open base and trains it with targeted data and rubrics, at a fraction of the cost of a closed flagship. If more teams validate that path, pricing for professional-services software will come under pressure.

Teams curious about it should run a few real documents from their own field, such as diligence files, contracts or research reports, through it and their current model, then compare quality and cost per task before deciding whether to switch.