Latest in AI
Ultramarine 1 Flash Weights Released: License, Hosting and API Pricing
Paper Instruments opens a new deployment path for its knowledge-work model, with important licensing and serving differences.
Paper Instruments has released Ultramarine 1 Flash's weights, expanding access beyond its existing API. The October 28 announcement adds a self-hosting path for the knowledge-work model, while Fireworks provides serverless inference and Baseten offers dedicated deployments. The license and hosting choices deserve as much attention as the benchmark claims.
What is new about this release
Ultramarine 1 Flash originally launched on October 14. This is a weights and deployment update, not a newly introduced model. Founder Daanish Khazi's announcement also says dedicated Fireworks deployments and fine-tuning access are forthcoming, rather than available through this launch.
The model targets professional tasks such as preparing documents and working through business information. The new deployment choices let teams evaluate those workflows in a different operating environment, rather than relying solely on the original hosted service.
Download access and self-hosting requirements
The official Hugging Face model card is public, but accessing the model files requires accepting its conditions and sharing Hugging Face account information. This is gated access, not an anonymous download.
The card describes 320 billion total parameters, 18 billion active parameters and a native 1,048,576-token context window. Its tested serving example uses SGLang on four B300 GPUs with a smaller 262,144-token context configuration. That example should not be read as a minimum hardware requirement or a promise that the full context has the same operating cost.
For a team evaluating private deployment, weights availability gives more control over the serving environment. It also transfers capacity planning, runtime maintenance and performance testing to that team.
The license has a provider restriction
Do not treat this as an unrestricted MIT release. The Ultramarine license allows free use for qualifying users that are not model providers, subject to its conditions. Organizations, including relevant affiliates, that monetize third-party inference or fine-tuning access need a separate agreement with Paper Instruments.
That distinction matters to an inference company considering adding the weights to its catalog. A public repository alone does not establish permission to sell access. Teams should review the actual license against their intended use rather than relying on a platform's short license label.
API pricing depends on the provider
The Paper Instruments product page and Fireworks endpoint currently list different token prices:
| Per million tokens | Paper Instruments API | Fireworks serverless |
|---|---|---|
| Input | $0.15 | $0.225 |
| Cached input | $0.03 | $0.045 |
| Output | $0.50 | $0.75 |
Fireworks identifies its endpoint as US-hosted and warns that serverless numerical precision can differ from training, dedicated deployments and published evaluations. These are reasons to test the actual endpoint you plan to use, not assume that every host reproduces a benchmark result exactly.
The useful next question is which deployment produces acceptable work at an acceptable total cost. That includes reviewing artifacts, recovering from failed tool calls and operating the infrastructure, not only buying the cheapest tokens.
I’m Paras. I work at the intersection of AI, engineering and distribution.
I help technical founders explain model releases and deployment tradeoffs in content developers can actually use.