Llama 3 and 3.1 made open weight the default baseline

Llama 3 and Llama 3.1 changed the default baseline conversation.

Before 2024, open weights felt like a special track. Llama 3 and 3.1 made them the thing you compare against by default.

That matters because open weights are not only about quality. They give you a real baseline, a deployment choice, and a way to decide what belongs in your own stack.

I think of the baseline as a test matrix instead of a slogan:

hosted frontier
open weight self-hosted
quantized self-hosted
small local fallback

Once the open route is strong enough, you have to explain why you skipped it. Maybe the latency budget is wrong. Maybe the operating cost is too high. Maybe the product needs a provider-managed update path. Maybe the local deployment cannot meet the hardware or privacy constraints. That is a better conversation than pretending hosted is the only serious option.

The shift is practical, not ideological. Llama 3 and 3.1 made open weights the default thing to test, which is the real change.

That baseline matters when you are deciding where a workflow lives. A routing service can compare hosted, open, quantized, and local fallback routes against the same examples. If the open route is good enough, the team has to explain why they are paying for the hosted one. If it is not, the failure is visible instead of assumed. That is a better starting point than treating open weights like a side quest.

In practice, that means the open route belongs in the first eval pass, not the optional one. The team can still choose hosted for support, policy, or capability reasons, but the choice has to be made after looking at the local baseline instead of before it.