Text Generation
Transformers
PyTorch
NeMo
English
nemotron
nvidia
llama-3

One of the few small models that does logical inference.

#10
by BingoBird - opened

My private testing method reveals minitron 4B to still be in the top tier of this size models, one year later Feb 2026.

It can make inferences that apply common knowledge not stated in the question context. Some of the winner models are:

2.4G SMART/Qwen3-4B-Hivemind-Inst-Hrtic-Ablit-Uncensored-Q4_K_M-imat.gguf
2.4G SMART/Qwen3-4B-Instruct-2507-Q4_K_M.gguf
2.7G SMART/Qwen3-4B-Thinking-2507-Q5_K_M.gguf
2.9G SMART/Minitron-4B-Base.i1-Q5_K_M.gguf
3.0G SMART/Gemmasutra-4B-v3a-Q6_K.gguf
3.4G SMART/Entropy-Qwen3-4B-Base.i1-Q6_K.gguf
4.0G SMART/Qwen3-4B-Claude-Sonnet-4-RS-DS-ST.Q8_0.gguf
4.0G SMART/Qwen3-4B-Hivemind-Inst-Hrtic-Ablit-Uncensored-Q8_0-imat.gguf
4.0G SMART/Qwen3-Instruct-4B-Kepler-186f.Q8_0.gguf
4.5G SMART/LFM2-8B-A1B-Q4_K_S.gguf
4.7G SMART/Mamba-Codestral-7B-v0.1-Q5_0.gguf
5.8G SMART/Apriel-1.6-15b-Thinker.i1-IQ3_XS.gguf
6.2G SMART/gemma-3-12b-it-norm-preserved-biprojected-abliterated.i1-IQ4_XS.gguf

Most models, including newly released ones under 10GB, fail my logical inference test.

Thank you and congratulations.

Sign up or log in to comment