Qwen 3.6 35B
A large-scale mixture-of-experts (MoE) language model in Alibaba's Qwen series with 35B parameters, suited to local deployment and agent tasks.
Timeline (last 90 days)
The voice stack consists of four core components: Qwen 3.6 35B (a mixture-of-experts model), Whisper, Breeze TTS2 (3.5B parameters), and Celery VAD.
Qwen 3.6 35B uses a MoE architecture, and when all expert weights are placed on the CPU with only a small amount of weights kept on the GPU, it can maintain a generation speed of about 40 tokens/second.
In the 12GB VRAM configuration, Breeze TTS2 requires about 8.5GB of VRAM (weights plus runtime overhead) and must run entirely on the GPU, while the LLM portion takes about nearly 4GB at a 128K context.
For users with 4×DGX Spark, a tiered deployment is recommended: the supervisor agent runs GLM 5.3 Flash on TP4, worker agents run Qwen 27B, and sub-agents run Qwen 3.6 35B.
All Facts (4)
The voice stack consists of four core components: Qwen 3.6 35B (a mixture-of-experts model), Whisper, Breeze TTS2 (3.5B parameters), and Celery VAD.
50%UnverifiedQwen 3.6 35B uses a MoE architecture, and when all expert weights are placed on the CPU with only a small amount of weights kept on the GPU, it can maintain a generation speed of about 40 tokens/second.
50%UnverifiedIn the 12GB VRAM configuration, Breeze TTS2 requires about 8.5GB of VRAM (weights plus runtime overhead) and must run entirely on the GPU, while the LLM portion takes about nearly 4GB at a 128K context.
50%UnverifiedFor users with 4×DGX Spark, a tiered deployment is recommended: the supervisor agent runs GLM 5.3 Flash on TP4, worker agents run Qwen 27B, and sub-agents run Qwen 3.6 35B.
50%