A Five-Person Team Fine-Tuned an Open Model Into a Claude Rival for $300,000
A five-person startup called Paper Instruments launched a model on October 14 that it says rivals Anthropic's Claude Fable 5 on professional document work at roughly one-sixtieth the cost, after a post-training run its CEO Daanish Khazi says cost about $300,000. The model, Ultramarine 1 Flash, isn't a new base model at all — it's a reinforcement-learning fine-tune of Z.ai's open-weight GLM-5.3 Flash, a 320-billion-parameter, 18-billion-active-parameter multimodal model already available for anyone to build on.
The pitch is narrow by design: Ultramarine is tuned specifically for finance, legal, consulting, and operations work, producing Excel workbooks, PowerPoint decks, and Word documents rather than general chat. Paper Instruments reports the model leads its own six-benchmark knowledge-work suite, including a proprietary JobBench score of 62.9 versus 57.4 for Claude Fable 5 and 55.1 for GPT-5.6 Sol, while using about 6.5% fewer output tokens than its GLM base. On the company's own cost accounting, a representative task runs $0.13 on Ultramarine against $8.19 on Claude Fable 5 and $6.96 on Claude Opus 5.5.
The fine-tuning economics
The headline number isn't the benchmark score — it's the $300,000 training bill. Where frontier labs spend hundreds of millions on pretraining runs, Paper Instruments' team applied LoRA-based reinforcement learning on top of an already-strong open checkpoint, grading outputs against rubrics with GPT-5.6 Terra as judge. That is a meaningfully different cost structure than building a foundation model from scratch, and it's becoming a recognizable pattern: specialized teams treat open-weight releases from Z.ai, Alibaba, and others as a substrate, then compete on narrow, well-defined task domains rather than general capability.
Taylor Lorenz, journalist, reacting on X: "Beautiful to see more high performing open-source models hit the market."
The open-source asterisk
That reaction captures the framing Paper Instruments wants — and also the gap in it. Ultramarine 1 Flash is being marketed as open-source, but no weights have shipped. The API is live; Khazi says the trained parameters will follow in about two weeks, putting an actual release near the end of October. Independent write-ups of the launch, including analysis from OrcaRouter (which serves GLM-5.3 Flash itself) and from independent AI commentator Nicolas Deville, both flag the same discrepancy: the company's benchmark claims are unaudited and self-graded, and until weights are public, 'open-source' describes an intention rather than an artifact. OrcaRouter's write-up also notes an internal inconsistency in Paper Instruments' own materials — the company says it leads four of six benchmarks, but its published table shows outright leads on only three, with Gemini 3.8 Flash and Claude Fable 5 still ahead on DiligenceBench and APEX-Agents.
What is unambiguously open is adjacent to the model rather than the model itself: Paper Instruments has published its Feather agent harness and forked versions of Python's Office-document libraries (paper-docx, paper-pptx, paper-xlsx) as public repositories, tools other developers can use regardless of what happens to Ultramarine's weights.
Why it matters beyond one launch
Ultramarine is a small story on its own, but it's a clean example of a larger shift this category has been tracking all year: the real competitive layer in open models is increasingly one step removed from the base-model labs. GLM-5.3 Flash, Qwen, and DeepSeek's open releases are becoming the raw material that smaller, better-capitalized-than-you'd-expect teams fine-tune into vertical products, undercutting frontier-lab pricing on specific workflows rather than trying to out-benchmark them generally. It also raises a labeling problem the open-source community hasn't settled: when a company calls a closed, API-only endpoint 'open-source' because weights are promised later, that promise is doing a lot of marketing work before it's done anything else. Whether Ultramarine's weights actually land in two weeks, and what license they carry when they do, will be the real test of the claim.