7 text
Xiaomi's open-source efficiency model: a 309B MoE with 15B active parameters per token and a hybrid attention mechanism.
Open-weight dense 27B vision-language model from Qwen for long-running agent tasks.
NVIDIA's open 30B MoE built for high-throughput agentic workloads.
Efficient Mixture-of-Experts Gemma 4 with only 4B active parameters per token.
Compact 9B vision-language model from the Qwen3.5 family.
Efficient vision-language MoE with linear attention and 3B active parameters.
OpenAI's compact open-weights model for efficient inference.