Key Assessment
Luo Fuli, head of Xiaomi’s large model team, gave a 3.5-hour technical interview in late April 2026—her first long-form public technical discussion since joining Xiaomi from Alibaba DAMO Academy and DeepSeek.
Core Viewpoints
1. Pre-train Gap Nearly Closed
Luo Fuli believes the gap between domestic top teams and Anthropic in pre-training is rapidly narrowing, and in some dimensions already closed.
| Dimension | Past | Present |
|---|---|---|
| Model Quality | International lead | Gap significantly narrowed |
| Training Methods | Insufficient experience | Methodologies converging |
| Compute Scale | Severely limited | Optimizations can compensate |
| Competition Focus | Pre-train scale | Agent RL |
2. Agent RL is Next Battleground
When pre-training is no longer a moat, competition shifts to Agent Reinforcement Learning:
- Real environment interaction: Agents must learn in real toolchains, not just synthetic data
- Multi-step decision making: From single-turn dialogue to multi-turn tool calling
- Self-correction: Can agents discover and fix errors autonomously
- Task decomposition: Planning and execution strategies for complex tasks
3. Talent Selection: Empty-Cup Mindset
Luo Fuli revealed her intern selection criteria—people with strong learning ability and curiosity:
People who can maintain an empty-cup mindset and think from first principles are rare. Strong learning ability gives them the power to quickly enter new roles.
From DeepSeek to Xiaomi: Technical Evolution
| Phase | Organization | Core Direction |
|---|---|---|
| Alibaba DAMO | Basic model pre-training | Early LLM exploration |
| DeepSeek | MoE + Open Source | MiMo series MoE architecture |
| Xiaomi | Edge-cloud + Agent | MiMo series + hardware ecosystem |
Industry Reflection on Claude Opus 4.6
Luo Fuli discussed the impact of Claude Opus 4.6 and similar 2026 technologies:
- Anthropic path: Building complete developer toolchain via Claude Code → Cowork → Agent Teams
- Domestic response: Cannot just follow; need differentiation in Agent RL and vertical scenarios
- Open vs. Closed: Open source community feedback speed is irreplicable advantage
Recommendations
| Role | Action |
|---|---|
| Model Developers | Make Agent RL core R&D direction; pre-train marginal returns diminishing |
| App Developers | Use MiMo Orbit free quota, low-cost Agent scenario testing |
| Job Seekers | Strengthen Agent framework and toolchain experience |
| Investors | Focus on teams with Agent RL capabilities and real-scenario data |