I can confirm Mistral 3 Small is an excellent model for many tasks and is an excellent base to Fine Tune on.
I’m currently using it in a number of places in the platform I’m building but 1 area on note is its use as the base for a set of LoRA adapters I use for initial Intent filtering and Classification.
For the above task specifically it’s been great. Not only has the accuracy been high across LoRA classifiers but the time to first token is amongst the best results I’ve seen.
Given its size, its speed and its accuracy for its size it’s a really appealing model and as I test more features across models I expect it to win in many use cases. It’s not going to be your core planner and of course there’s specific models for embedding, reranking, etc but in a lot of places it’ll be hard to beat the combination of features this model offers.
imo it is one of the best things Mistral has produced and isn’t getting enough love in the space so I’m happy you highlighted off the top.
Hi Ronan,
I can confirm Mistral 3 Small is an excellent model for many tasks and is an excellent base to Fine Tune on.
I’m currently using it in a number of places in the platform I’m building but 1 area on note is its use as the base for a set of LoRA adapters I use for initial Intent filtering and Classification.
For the above task specifically it’s been great. Not only has the accuracy been high across LoRA classifiers but the time to first token is amongst the best results I’ve seen.
Given its size, its speed and its accuracy for its size it’s a really appealing model and as I test more features across models I expect it to win in many use cases. It’s not going to be your core planner and of course there’s specific models for embedding, reranking, etc but in a lot of places it’ll be hard to beat the combination of features this model offers.
imo it is one of the best things Mistral has produced and isn’t getting enough love in the space so I’m happy you highlighted off the top.
Cheers,
Christopher (@communicating on twitter).
Thanks