Why Xiaomi's MiMo Improved So Fast: Autonomous Training, Open-Source Misconceptions, and Talent Value Assessment

Correcting an impression: obscurity does not mean inactivity

Xiaomi released MiMo-V2.6-Pro with an official note dated September 22, 2026, introducing Pro and Flash native multimodal models and opening weights and research resources. However, the technical trail begins earlier: a MiMo technical paper submitted in May 2025 documented MiMo-7B pre-training on 25 trillion tokens and post-training using verifiable math and coding tasks. The official code repository also published corresponding model resources. This record shows Xiaomi already possessed public technical accumulation for training base models, though the 7B record alone cannot automatically prove the flagship V2.6-Pro's full provenance; the specific model's technical report, weights, and training notes must be examined.

What "using open-source models" actually means

Discussions of autonomous R&D often conflate four distinct activities:

Using open-source training frameworks — frameworks help schedule compute, collect trajectories, and optimize parameters; two companies using the same framework can train models with different origins.

Adopting public architectures or research methods — MoE, attention mechanisms, RL algorithms are technical routes; similar routes do not imply inheriting the same model weights.

Starting from existing model weights for continued pre-training or fine-tuning — the base source must then be explicitly disclosed.

Distillation — a student model learns from a teacher's outputs, reasoning traces, or other signals; the student's initialization source and the teacher's source are separate questions.

Xiaomi simultaneously released a model named MiMo-V2.6-Distill-Qwen-9B. Its independent model card should be checked, but one distillation variant in the same batch does not prove Pro uses the same base. Official listings treat them separately. Therefore, "is it a wrapper" should be replaced with verifiable questions: where do initialization weights come from? Which training stages did the team complete? How transparent is the training process? Are licenses and sources clear? Model self-identification or stylistic similarity are only clues; similar answers may stem from similar corpora, prompts, or distillation. Even similar architecture and tokenizer require further evidence to establish weight inheritance.

Behind the six-day training lies longer investment

Release materials highlight a figure: Pro's public RL training phase took under six days at a cost of approximately $2.62 million. This number is scoped to a reinforcement-learning stage only and cannot represent the total from-scratch R&D account. Prior base-model training, data construction, failed experiments, engineering development, and team investment must be counted separately. An accompanying diagram illustrates the post-training feedback-optimization loop (orange cycle) positioned after the base model, clarifying why local training cost cannot substitute for full R&D cost.

Official training curves on DeepSWE v1.1 show the orange Pro curve rising from 58.4 to 72.6 with intermediate fluctuations, ending above the starting point. This demonstrates performance change under that specific benchmark and its conditions, but cannot be extrapolated to "all tasks improved equally" nor to "talent investment has paid back." The engineering capability to produce such change is more telling: complex Agent training requires model-environment interaction, trajectory logging, result judgment, and feedback-driven optimization. Unstable environments, flawed reward signals, or train-inference mismatch can amplify errors despite more compute. This explains why model competition depends not only on parameter scale but also on building trustworthy feedback, diagnosing failures, and continuously improving the training process.

How to evaluate Luo Fuli's addition

Luo Fuli publicly confirmed joining Xiaomi's MiMo team in November 2025. Reports of a "ten-million-yuan annual salary" remain unconfirmed; insiders suggest the figure may be exaggerated. By December 2025 she appeared as MiMo LLM lead at Xiaomi's partner conference. Placing these facts alongside the team's sustained delivery allows discussion of potential talent value, but "progress after joining" does not quantify how much progress one person caused. Public announcement time ≠ actual work start; model projects involve multi-person, multi-phase collaboration; release dates ≠ R&D start dates; compute budgets, existing team, and tooling may change simultaneously. We observe actual output but lack a counterfactual — what the same Xiaomi team with the same resources would have delivered without this hire — limiting causal attribution.

Nevertheless, inability to precisely attribute does not preclude analyzing a technical lead's role. In high-cost R&D, a mature lead can influence team efficiency by choosing experiment directions, spotting anomalous results, adjusting evaluation methods, and organizing research-engineering collaboration. An earlier-caught training issue may reduce later wasted spend; a more reliable experimental process may enable more researchers to sustain output. These are mechanism analyses of how talent creates value, not a checklist of Luo Fuli's personal achievements. Public information supports discussing the mechanism, but not itemizing her contribution.

Technical progress and commercial return must be accounted separately

The article proposes three evaluation layers:

Technical Output

What to check: Models, training records, reproducible evaluations

What it answers: Did the team deliver new capabilities?

Organizational Capability

What to check: Experiment efficiency, collaboration, stable iteration mechanisms

What it answers: Can it continuously produce technical results?

Commercial Return

What to check: Revenue, cost savings, deployment and maintenance costs

What it answers: Did investment generate economic value?

Xiaomi's public materials make technical output concretely discussable. Organizational capability requires longer observation. Commercial return needs attributable business data; release frequency and leaderboard scores alone cannot compute it. Even if a strong lead helps avoid a major training failure, one must know the avoided loss and the resources spent to estimate gain. If the model never reaches production, technical progress may not translate to revenue or cost reduction.

For developers, the same distinction applies: judge a model by real tasks — let it fix the same code issues under identical tool permissions and context, then examine test results, tool-call errors, total latency, and human rework. API price is only part of cost; wrong results also carry cost.

Focus on whether it can keep improving

MiMo's public record is sufficient to discard the inference "Xiaomi was unknown, therefore it had no autonomous R&D." Meanwhile, open-source reuse and the existence of a distillation variant must be examined per object, not broadly extrapolated to the whole model family. The progress after a researcher like Luo Fuli joins is worth watching. A more explanatory lens is how mature experience combines with the existing team, compute, and experimental facilities to form a sustained delivery capability. This remains analysis based on public information; personal contribution and commercial return cannot yet be quantified.

Going forward, the key signals are whether training resources can be externally audited, whether the model works stably on real tasks, and whether subsequent iterations continuously improve quality and total cost. These outcomes, more than the "hired a genius" label, will show what talent and organizational investment ultimately produced.