Lenovo has opened pre-orders for the P7 AI host, featuring a heterogeneous compute setup with Cixi's P1 SoC and Houmo's M50 dNPU. The device delivers 190 TOPS of local AI power, supporting 122B parameter LLMs with a 128K context window at just 30W total system power.
Tokens Per Second
TPS
A metric indicating how many text units an LLM generates per second.
Curated Intelligence
OpenAI's Codex lead launched a 28-day intensive update sprint, delivering a ~50% inference speed boost for GPT-6 Astra and Sol on day one. Despite the performance optimization, community reaction was largely sarcastic, viewing it as PR damage control following prior latency issues and commercialization controversies.
The vLLM community has optimized serving for Qwen3.5's hybrid attention architecture (combining full-attention and GDN layers), achieving over 25K TPS per GPU on NVIDIA GB200 NVL72 systems. By addressing heterogeneous state transfer and accelerating GDN computation on Blackwell GPUs, this solution significantly boosts the performance ceiling of disaggregated serving.