Open source is a legacy of the 1980s free software movement, built on the principle that users should have the freedom to run, study, modify, and distribute software. In the late 1990s, developers associated with the Linux operating system and the Netscape web browser promoted the term “open source” to refer to these ideals.
The Divide Between Open Source and Open Weight
As the movement evolved, organizations developed open-source licenses to specify how source code could be used and distributed. These include the MIT License, Apache License, Berkeley Software Distribution, and the Gnu General Public License, with each specifying potential restrictions on how software patents applied to the source code. The Open Source Initiative’s definition of a fully open-source AI model specifically includes training data as a key element, though some developers question the feasibility of distributing such enormous datasets.
In this scenario, developers receive the trained model—the “weights” or encoded knowledge learned during training—which they can download and run on their own infrastructure. However, the code and training data used to produce that model may remain secret.
Publishing the code is the most direct way to build toward a robust and reliable harness. You can read the source to see exactly how it works, from context assembly to tool call dispatch.
SpaceXAI
SpaceXAI further explained that open sourcing the code for Grok Build makes the harness easier to explore and extend for those working with subagents, hooks, plugins, skills, or MCP servers, as the source serves as the definitive reference for how each is loaded and invoked.
Meta released LLaMa on Feb. 24, 2023, providing both the weights and the “inferencesource code—the instructions that run the model. Yet, open-source organizations such as the Open Source Initiative argued that LLaMa's licensing guidelines prohibited commercial reuse, meaning it did not meet the strict definition ofopen source.”
The Economic Engine of Open Weight Infrastructure
In July, Fireworks announced a $1.505 billion Series D funding round at a $17.5 billion valuation. During the same period, Together AI secured $800 million, and Baseten raised approximately $1.5 billion.
These companies do not build their own foundation models.
Continuous batching and cache management to handle high traffic.
Quantisation and autoscaling to optimize hardware use.
Latency guarantees and integrated billing systems.
Fireworks reported that more than 95% of its served tokens come from models specialized using customer data, such as insurance claims or support records.
Strategic Logic: Democratization vs. Control
These are our most advanced open weight reasoning models… They give anyone, from individual developers and local nonprofits, to large enterprises and governments, the freedom to run and customize AI on their own infrastructure, democratizing access to AI across industries, communities, and countries globally.
Open-Source AI vs Open-Weight AI Explained | What’s the Real Difference?
OpenAI
OpenAI further argued in a submission to the White House Office of Science and Technology Policy that the debate between open and closed source is a false choice, suggesting that both can work complementarily to ensure AI is built on American rails.
