# An economy for open weight AI models > Source: > Published: 2026-10-29 17:12:48+00:00 Today we're releasing the weights of Ultramarine 1 Flash, and we wanted to share a few notes on the economics of open-weight models. Traditionally, tokens have been delivered as a final good, bundled by the developers of the weights. Frontier labs train models and serve those models on compute they provision. IBM similarly bundled software, compute, and packaging as mainframes to enterprise customers who could afford it. Personal computing, which eventually replaced this model, was successful, in part, due to how it unbundled and separated concerns. Microsoft licensed software to Dell who built a scaffold around Intel chips. We are noticing a similar value chain emerging for weights. Model developers have started licensing weights to inference providers who build inference software on compute designed by chipmakers and operated by a neocloud. Thus far, this model has not seen maturation for open-weight developers. While closed-weight developers, neoclouds, inference providers, deployment companies, and application companies have created incredible demand and obtained the scale required to command margins, training a frontier class model and distributing it for free has historically been unprofitable. In late 2026, two things are different: (1) post-training is quickly becoming cheaper as RL recipes, base models, and infrastructure improve, and (2) new licensing structures allow open-weight developers to capture value at zero marginal cost. Inference and pre-training each require large, talent-dense organizations that deploy significant CapEx to compete on optimization, scale, and generality. In traditional software, you don't need massive capex to be market competitive, and this is why startups with good product sense have historically fared well. Our bet is that post-training is more open-ended and lends itself to a business model that is closer to traditional software, where taste, judgment, and small teams will produce better products than legacy incumbents. Across a variety of domains spanning knowledge work, search, design, and so on, we expect this new working model of intense data-training codesign to outperform the data vendor outsourcing at frontier labs. As such, we expect to see a rich ecosystem of open-weight developers post-training domain-specific models that are Pareto optimal for the large chunks of GDP that have thus far been untouched by AI. These open-weight developers will benefit from open-source software, deploy on existing inference provider networks, and build upon a growing catalog of open base models. Open-weights will be an essential component of each enterprise’s intelligence stack. We’re releasing Ultramarine 1 Flash under a simple license that’s free and open to use for any person or business directly using the model, internally or commercially. If you are an inference provider selling access to the model, we are licensing the model under straightforward terms. As compute demands grow at a fast clip and hardware becomes increasingly heterogeneous, we expect many more kinds of inference providers, and we are excited to follow up with a self-serve agreement to serve our weights. Today, our training partner, Fireworks, is serving Ultramarine 1 Flash on serverless. Serverless is on a shared pool and is a good choice if you want to pay-by-token and optimize for speed. For higher precision options and more control, dedicated deployments will be available on Fireworks shortly, in addition to the ability to fine-tune Ultramarine on Fireworks’ Training API. We’re excited to deepen our partnership with Baseten, who is serving the model on dedicated deployments. The Baseten team worked closely with us to tune the model for latency sensitive workloads. We have several more providers and routers coming online in the coming weeks and are excited to announce different ways of accessing and tuning the model to fit your needs.