Memo VR
Memo VR is the model series discussed by Luo Fuli / 罗福莉 in 138. 对罗福莉3.5小时访谈：AI范式已然巨变！OpenClaw、Agent范式很吃后训练、卡的分配、组织平权. The source presents it as Xiaomi’s model line for agent-era workloads, including Pro, Omni, TTS, Flash, long-context efficiency, multimodal perception, and voice output.
The episode describes a role-based architecture. Pro is associated with understanding, cognition, and complex scheduling; Omni is associated with video, audio, image, and text perception; and TTS handles voice output and expressive speech. Luo says the models are not collapsed into one model mainly because speed, cost, and price still matter in real agent workflows.
Key Claims
- Memo VR is positioned around Agent Harness scenarios, where models need long context, tool interaction, fast generation, and cost-aware orchestration.
- Flash and Pro are described as related structures trained around similar timing, with choices such as hybrid attention, sliding windows, MTP, and KV-cache tradeoffs made for long-context efficiency.
- The source says Pro encountered harder large-model training issues such as numerical instability, loss spikes, and expert-distribution anomalies.
- Multimodal work is treated as useful for agent action and world knowledge, but the source does not claim that multimodal training alone guarantees AGI.
- TTS is framed as a large-scale generalization problem where a simpler architecture plus scale can follow nuanced natural-language descriptions of emotion and rhythm.
Connections
- Luo Fuli / 罗福莉 and Xiaomi — team and company context.
- Agent-Optimized Model Architecture — architecture choices shaped by agent workloads.
- Agent Post-Training and Agent RL — post-training and rollout infrastructure that the source says agent models increasingly need.
- Open Claw, Open Cloud, and Model Harness Co-Evolution — framework context for the model-series strategy.
- Frontier Model Scaling, Long-Horizon AI, and Model Workflow Fit — scale, context, and practical model-selection themes.