Ultramarine 1 Flash is Fable-class at professional work for one sixtieth of the cost

What Ultramarine 1 Flash is

Ultramarine 1 Flash is a multimodal agentic model that keeps GLM-5.3 Flash's architecture and FP8 quantization scheme, with a 1M-token context window. Paper Instruments, a company building AI for professional services, post-trained the base with large-scale reinforcement learning on synthetic professional tasks rather than on code.

The release is unusually complete for an open model:

• Weights on Hugging Face under a custom Ultramarine 1 Flash License.
• The agent harness, training code and evaluation code open-sourced as feather-train.
• Paper Office, a set of document tools (paper-docx, paper-pptx, paper-xlsx) for safer office-file editing inside agent harnesses.
• A hosted API priced identically to GLM-5.3 Flash: $0.15 per million input tokens, $0.03 cached, $0.50 output. Baseten also lists the model for managed inference.

The benchmarks and the cost claim

The headline numbers cover six knowledge-work benchmarks. Ultramarine 1 Flash leads the field on JobBench (62.9, ahead of Gemini 3.8 Flash's 60.6 and Claude Fable 5's 57.4), on the internal PI Knowledge Work suite (59.3 vs Fable 5's 57.8), and on FrontierFinance (53.3). It places second on AutomationBench behind DeepSeek V4.1 Flash and third on DiligenceBench and APEX-Agents. Across all six it averages 4.1 points above its GLM-5.3 Flash base while generating 6.5 percent fewer tokens, reasoning included.

The cost argument is the sharper one. On PI Knowledge Work, a suite that asks models to deliver finished Excel workbooks, PowerPoint decks and Word documents inside live business systems, Ultramarine costs $0.13 per completed task. Claude Fable 5 at maximum reasoning costs $8.19 per task and scores lower; GPT-5.6 Sol costs $2.09 and scores lower; Claude Opus 5.5 ($6.96) and GPT-6 Sol ($1.26) stay ahead but only by about three points. Paper Instruments says PI Knowledge Work scores correlate with the Artificial Analysis Intelligence Index at r = 0.81.

The caveats: the headline suite is internal, every model ran in the same harness at official API pricing, and the public benchmarks are graded with rubrics rather than unit tests. Treat the absolute numbers as directional; the cost gap is the part that survives scrutiny, because the API price and token counts are public.

The research contribution: RL on work you cannot unit-test

Most agentic post-training so far has concentrated on verifiable tasks, where coding dominates, because code gives you a reward signal for free: the tests pass or they do not. Professional work does not. A diligence memo or a litigation chronology cannot be graded by a regex, and its quality turns on judgment.

Paper Instruments' answer is task and rubric co-design. Each training task is packaged under the Harbor specification: an instruction written the way a professional would receive it, a runnable environment with a Dockerfile and live services, and a weighted rubric. Each rubric criterion draws a binary boundary over possible answers, stating the substantive condition that earns credit with the facts, thresholds and context needed to check it, so a grader model can mark it met or unmet without requiring one golden solution. The technical report walks through a board-forecast task where the criterion separates what the board can approve now from what still needs counsel clearance, and shows how an overfit or vague criterion corrupts the reward.

The curriculum is fully synthetic and combinatorial: four domains, 19 practice areas, 88 environment types, with each task sampled across axes like deliverable format, stakeholders, source conflicts and deal stage, giving a configuration space of 10 to the 7th through 10 to the 10th combinations per practice area. That breadth is deliberate; it is what keeps LLM-generated training data from collapsing into one mode. The training run itself is modest by frontier standards: LoRA rank 64, CISPO loss, a 131K-token training context, up to 250 agent turns per episode.

How to run it

Self-hosting is realistic for a well-equipped team: Paper Instruments serves it with SGLang on four B300 GPUs, and the Hugging Face card has the exact launch command, including the two non-default flags for the BF16 shared experts and the full context length. Recommended sampling is temperature 1.0 and top_p 0.95, and the card advises running it inside an agent harness rather than as a chat model. If four Blackwells are not in your budget, the API at $0.15 per million input tokens or Baseten's managed listing get you the same weights.

What Ultramarine 1 Flash means

Two signals. First, the RL frontier is moving from verifiable to judgment work. Code was the beachhead because verification was free; this release shows rubric-graded synthetic tasks can carry the same trick into finance, law and operations, which is where most enterprise agent budget actually sits.

Second, the open-weights cost curve keeps collapsing onto specific professional verticals. A post-trained GLM-5.3 Flash variant now beats a frontier model on knowledge-work benchmarks at one sixtieth of the per-task cost, in the same week Meituan's LongCat-2.5 put a 1.6-trillion-parameter coding agent model on an Anthropic-compatible API at $0.30 per million tokens and GPT-6.1 Terra halved computer-use pricing. If you build agent products, routing routine professional work to models like this and reserving frontier spend for the hard tail is no longer an optimization. It is the baseline.

FAQ

Sources

1. 01Model page and benchmarks: Ultramarine 1 Flash - Paper Instruments (opens in a new tab)
2. 02Technical report: Ultramarine 1 Flash: Scaling Non-Verifiable RL - Paper Instruments (opens in a new tab)
3. 03Weights and serving: paperinstruments/Ultramarine-1-Flash - Hugging Face (opens in a new tab)

Elsewhere on this site

The selected work catalogue lists the SaaS products and dashboards behind notes like this one.