[ivorbuk]

We use llama models as the comparative cost, at the time of decision making (9 months ago?), was cheaper++ then closed source models and comparative to other models available through bedrock (we use via aws bedrock).

Strong reasoning at the time was also to, at some point, take this on-prem/self-host - for privacy. More a comfort blanket for some of our customers/partners and a future requirement rather than a right now thing.

From a capability perspective it's everything we need - tho we are not taxing or pushing any boundaries... Our use cases are mainly processing/structuring/summarising incoming text and we run a few agents doing a variety of stuff. We have a bit of technical jargon (motorsports engineering related) and we saw good performance with llama over our previous use of Mistral/Claude/OpenAI.