Frontier-class open weights push value from models to applications

Meta's 405 billion parameter model comes with a license that permits distillation. Top-tier capability becomes a cheap input, and the edge moves to data.
Anyone can now download the weights of a 405 billion parameter model and use its outputs to train other models. Meta released it on 23 July as part of Llama 3.1 and calls it the first frontier-level open model. It was trained on more than 16,000 Nvidia H100 GPUs, and on Meta's own evaluations it is competitive with GPT-4o and Claude 3.5 Sonnet.
Value is leaving the model layer. For two years the most capable models sat behind APIs controlled by a few labs, and those labs set the price of premium reasoning. A frontier-class model that any company can download, host and fine-tune caps what a closed provider can charge for comparable work, so the margin on general model access narrows. Value moves to the compute that serves models and to the companies that turn them into products with their own data.
The price pressure comes from both ends of the market. Meta puts the cost of running the 405B model on one's own infrastructure at roughly half that of closed models like GPT-4o, and Amazon Web Services, Microsoft Azure, Google Cloud, Nvidia, Databricks and Groq offered it from the first day. Meta can give the model away because it earns its money elsewhere. Five days before the release OpenAI launched GPT-4o mini at 15 cents per million input tokens, more than 60% cheaper than GPT-3.5 Turbo, so in one week both the top and the bottom of the model market got cheaper.
The license matters as much as the model. Teams can now use the largest open model to generate synthetic training data and distill its skills into small models that are cheap to serve and tuned to one task. A startup can also run its own model inside a customer's environment, which answers data-control objections in finance, health and the public sector. The catch is operational: even in 8-bit form the 405B model needs a full server node of accelerators, so most teams will use it as a teacher and ship something smaller. Model choice becomes a cost decision to revisit every quarter, and the lasting advantage has to be built in proprietary data, workflow and distribution.
Closed labs still have room to answer, with further price cuts or by moving up to capabilities that open models do not yet match. The price per token that hosting providers settle on for the 405B model will show how far open weights pull down the market rate for this level of capability. How many enterprises move production workloads from closed APIs to self-hosted open models will show whether the cap holds in practice.
Frontier capability is turning into an input that several suppliers compete to provide, as image generation did once open weights arrived. Teams whose value sits in data, workflow and customer relationships gain from every price cut, and Fanaura's mandate in AI-native companies favours them over anyone whose edge is owning the model.
Building an AI-native company? Apply in five minutes. →
More from Fanaura
Building on open models with your own data and distillation pipeline? Send us the project.
Fanaura invests its own capital and the capital of its limited partners (LPs), usually through a separate SPV for each deal. If we see a fit, we get in touch.
Get the news by email
One short note each weekday. No marketing, and you can unsubscribe at any time.