A 70-billion model matching a 405-billion one. Efficiency is the quiet trend of 2024
Meta released Llama 3.3 70B on 6 December 2024, claiming performance near its far larger predecessor at a fraction of the serving cost.
On 6 December 2024 Meta released Llama 3.3 70B, saying it delivers performance similar to the 405-billion-parameter Llama 3.1 at a fraction of the cost to serve. The improvements centre on instruction following and multilingual capability.
Set aside whether the claim holds in every case. The direction has been consistent all year, and it is the trend with the most practical consequence for an SME.
Why smaller-but-equal matters more than bigger
It shows up in your invoice. A model that is six times smaller to run is a model that costs less per answer. Through 2024, the price of a given level of capability fell repeatedly, and this is one of the mechanisms.
It changes where a model can live. A 70-billion-parameter model is demanding but tractable on real hardware. A 405-billion one is a data-centre proposition. That difference decides whether self-hosting is a conversation at all.
It makes speed possible. Smaller models answer faster, and for anything a person waits for — a lookup during a customer call, a check while writing a quotation — response time matters more than the last few points of accuracy.
The lesson for choosing tools
Parameter counts are marketing. They were a rough proxy for capability in 2023 and they are not one now. A vendor who leads with the size of the model underneath is telling you about their supplier, not about your outcome.
The useful test has not changed: give the tool ten real examples of your own work — your quotations, your customer emails, your German, your abbreviations — and see what comes back. That takes an afternoon and tells you more than any benchmark table.
What to expect from here
Assume that whatever capability you are paying premium prices for today will be available from a smaller, cheaper model within a year. That is not a reason to wait, because the work of identifying which task to apply it to takes longer than the wait would save.
It is a reason to keep contracts short and to avoid building anything that only works with one specific model.