Xiaomi’s AI model lead: As AI competition shifts to the Agent era, self-evolution is a key event on the path to AGI

Xiaomi’s large-model team lead, Luo Fuli, held an in-depth interview on the Bilibili platform on April 24 (video ID: BV1iVoVBgERD). The interview lasted 3.5 hours, and it was her first time, as a technical leader, to publicly and systematically lay out her technical viewpoints. Luo Fuli said that the large-model competition track has shifted from the Chat era to the Agent era, and she pointed out that “self-evolution” will be the key event for AGI in the coming year.

From the Chat Era to the Agent Era: Core Technical Judgment

（Source: Bilibili）

Based on Luo Fuli’s statements in her Bilibili interview, she said that the focus of the 2026 large-model competition has shifted from general dialogue quality to continuous autonomous execution capability in complex tasks. In the interview, she said that current top models can autonomously optimize in specific tasks and continue to execute steadily for 2 to 3 days without human intervention or adjustment. She emphasized that the breakthrough in “self-evolution” capability means AI systems have started to possess the ability to self-correct, and she singled out how Anthropic’s technical roadmap and technical variables such as Claude Opus 4.6 will affect the entire AI ecosystem.

Xiaomi’s Compute Allocation Adjustment and Pre-train Generation Gap Assessment

According to what Luo Fuli disclosed in the interview, Xiaomi has already made major adjustments to its compute allocation strategy. She explained that the compute allocation ratio commonly used in the industry is Pre-train:Post-train:Inference = 3:5:1, while Xiaomi’s current strategy has been adjusted to 3:1:1, significantly compressing the proportion allocated to post-training while simultaneously increasing resource investment in the inference stage.

In the interview, she explained that this shift is driven by the maturity of the Agent RL Scaling strategy, meaning that post-training no longer requires piling up large amounts of compute, and the increased resources on the inference side reflect the need for an Agent’s on-the-ground scenarios to have real-time responsiveness.

Regarding the pre-train generation gap issue for domestic large models, Luo Fuli said in the interview that this gap has been shortened from the past 3 years to several months, and that the current strategic focus is moving toward Agent RL Scaling. Luo Fuli’s career history includes Alibaba’s DAMO Academy, Huanfang Quant, and DeepSeek (core developer of DeepSeek-V2). She joined Xiaomi in November 2025.

MiMo-V2 Series Technical Specifications and Open-Source Rankings

According to a MiMo-V2 series announcement released by Xiaomi on March 19, 2026, this time, three models were released in one go:

MiMo-V2-Pro: Total parameters of 兆; activated parameters 42B; hybrid attention architecture; supports million-level context; task completion rate 81%

MiMo-V2-Omni: Multi-modal Agent scenarios

MiMo-V2-TTS: Speech synthesis scenarios

According to the announcement, the open-sourced MiMo-V2-Flash ranked second on the global open-source model leaderboard, with an inference speed reaching 3x that of DeepSeek-V3.2.

Frequently Asked Questions

How does Luo Fuli define “self-evolution,” and why does she think it is the most critical event for AGI?

According to Luo Fuli’s interview statement on Bilibili on April 24, 2026 (BV1iVoVBgERD), she said that currently, top models can autonomously optimize in specific tasks and execute steadily for 2 to 3 days without human intervention, and that she characterizes “self-evolution” as the most critical event for AGI development in the coming year.

What specific adjustments has Xiaomi made to its compute allocation ratios, and what is the rationale behind it?

Based on Luo Fuli’s disclosures in the interview, Xiaomi’s compute allocation ratio has been adjusted from the industry-standard Pre-train:Post-train:Inference = 3:5:1 to 3:1:1, significantly compressing the proportion for post-training. She explained that this adjustment is due to improved post-training efficiency after the Agent RL Scaling strategy matures, as well as the need for the Agent’s deployed scenarios to have real-time responsiveness on the inference side.

What is the open-source ranking and speed performance of MiMo-V2-Flash?

According to Xiaomi’s official announcement released on March 19, 2026, the open-sourced MiMo-V2-Flash ranks second on the global open-source model leaderboard, with inference speed 3 times that of DeepSeek-V3.2. The task completion rate of the flagship MiMo-V2-Pro is 81%.