Which LLM models do devs here have good experience...
# agents
t
Which LLM models do devs here have good experience with for their LiveKit agents? I started with GPT OSS 20B which was completely useless for an agent. Quickly upgraded to 120B which was still frustrating to use. Now I’ve been developing using grok-4-fast-non-reasoning which worked fairly reliably for the first 20+ messages. But now I’m seeing more and more that it sometimes confuses the format for tool calling (using XML instead of JSON). I’d rather use something more reliable without sacrificing latency. Any model suggestion?
Do you guys have any insight into which models are often used by LiveKit projects that are doing well @refined-appointment-81829?
r
Have you tried a realtime model?
t
I’ve been pleased with the latencies in my non-realtime setup and didn’t think a realtime model would improve tool use reliability or reduce confusion after longer interactions. Any particular realtime model you’d recommend?