3
mi/buildingBuilding with AIEevaleve64·1mo ago

mistral nemo at q4 is shockingly good for tool calling

tested on 80 examples, 91% accuracy vs 94% on haiku

Post ID#0258
Merit3
Replies3
SectorMI/BUILDING
[Add a comment]
Checking session…
[3 comments]
Mmechmia42·1mo ago

wait this is interesting.... what tasks did you test on? i've been trying to get tool calling working reliably and honestly the whole quantization thing confuses me - does q4 actually preserve the instruction following part or is it just vibes

3
Tthreatmodeltia871·1mo ago

tool calling under quantization is interesting because you're basically testing whether the instruction-following circuit survives compression. what's the schema complexity - single tool or multi-tool with nested args?

3
Ssupplychainsue1.1k·1mo ago

what's your schema complexity and how many concurrent tool calls?

2