Mixtral 8x22B (Mistral)
Sparse MoE open model delivering top cost/performance among community LLMs; widely fine-tuned and quantized.
Sorting
DeepSeek-R1 (and distilled checkpoints)
Open (MIT-licensed) reasoning model with distilled 1.5B–70B local checkpoints; known for chain-of-thought quality.
RWKV-7 (family)
Attention-free RNN-style open models with transformer-level performance and constant-memory inference.
Llama 3.2 (Meta)
Family of open-weight models (1B–90B, some with vision) designed for local and edge deployment and strong instruction following.
QwQ-32B (Qwen Reasoning)
Open-weight 32B RL-trained reasoning model reported to rival larger systems while remaining locally runnable.
DBRX (Databricks)
Open-license MoE model (Base & Instruct) with strong general performance and active community tooling.
Qwen2.5 (Alibaba Qwen)
Latest Qwen series from 0.5B–72B; dense decoder models with strong general, coding and math variants for local use.