Xiaomi MiMo: anatomy of a reasoning pipeline, from 7B to 1T

A technical reading of pretraining, MoE routing, distillation, and agent memory.

There is a reason to look closely at Xiaomi MiMo beyond parameter counts: the team releases recent models and publishes the technical work that accompanies them. For Kuaapy, this openness has value in itself. Being able to study a model, adapt its execution, and confront results broadens who can participate in AI development.

The V2.6 generation is part of that choice. The official announcement presents the release of the series, the technical report, the training environments, and the RL code; public cards provide access to checkpoints. This combination creates investigation opportunities beyond the company's hosted service. V2.6 release [1], official models [2].

An agent can get the answer right and waste dozens of calls reaching it. It can generate correct code and lose the constraint that should guide the next change. It can have a huge context and still manage task state poorly.

To analyze such a system, we need to follow three layers: the weights that define the model's policy, the infrastructure that trains and runs that policy, and the runtime that organizes tools and continuity. MiMo's public trajectory allows us to examine all three.

This analysis uses primary sources consulted on October 23, 2026. Metrics attributed to Xiaomi are results published by the authors; they do not represent independent reproduction by Kuaapy.

1. Reasoning starts before RL

The MiMo-7B report describes joint optimization of pretraining and post-training. The base received approximately 25 trillion tokens, a three-stage data mixture, and an auxiliary Multi-Token Prediction objective, or MTP. Data preparation includes mathematical content, code, and synthetic reasoning.

In post-training, the authors use SFT and reinforcement learning with about 130,000 verifiable mathematics and programming problems. The code reward considers test difficulty; resampling seeks to stabilize training. This is more specific than saying the model received a generic reward for every reasoning step.

The appropriate conclusion is that data, pretraining objectives, and post-training contribute at different stages. The paper does not support the claim that reasoning is exclusively a product of RL. Nor does an isolated correct final result distinguish generalization from memorization. That distinction requires its own evaluation design. MiMo-7B report [3].

2. Total scale and active computation are different numbers

MiMo-V2-Flash combines 309 billion total parameters with 15 billion active per token. Its report describes local attention with a 128-token window interleaved with global attention, at a 5:1 ratio. The 32K pretraining context was extended to 256K. V2-Flash report [4].

In the official cards for the V2.6 generation, the separation looks like this:

V2.6-Pro-RL — Total parameters: 1.02 trillion; Active per token: 42 billion; Routed / activated experts: 384 / 8; Declared context: 1 million

V2.6-Flash-RL — Total parameters: 309 billion; Active per token: 15 billion; Routed / activated experts: 256 / 8; Declared context: 1 million

Both receive text, images, video, and audio. They are distinct models, not merely two speed settings for the same checkpoint. Pro [5], Flash [6].

MoE reduces the set of experts used for each token. This does not make the entire model fit in memory corresponding to its active parameters. As an illustrative calculation, one trillion parameters at one byte per parameter already represents approximately one terabyte, before quantization scales, caches, and buffers. Distribution, offload, and weight transfer remain architectural questions.

Likewise, a declared maximum context does not by itself tell us latency, cost, or the quality of retrieving distant information.

3. When rollout and training choose different experts

During RL, one engine generates trajectories and another computes updates. In an MoE, different routing decisions can make these stages activate different experts for the same sequence. The problem appears in the relationship between the policy that produced the sample and the policy used to update it.

Rollout Routing Replay, or R3, records expert selections during rollout and reapplies them during training. The implementation keeps computing router probabilities in training, preserving the gradient path. The goal is to reduce divergence between engines without simply discarding difficult samples.

The paper's experiments include Qwen3-30B-A3B; their numbers should not be presented as MiMo measurements. The relevant result for this discussion is the mechanism: router behavior and consistency across stacks participate in RL stability. R3 paper [7].

4. MOPD integrates experts into the student's trajectories

Training one mathematics teacher and another software-engineering teacher does not automatically solve how to combine their skills. In MOPD, specialized teachers are developed by domain. The student produces its own rollouts and receives supervision from the corresponding teacher, with a dense per-token signal.

One formulation uses reverse KL divergence between student and teacher distributions on contexts visited by the student. The learned object is a joint policy; the procedure is not concatenating answers or simply averaging the teachers' weights.

The study compares the technique with alternatives on Qwen3-30B-A3B and reports industrial application in V2-Flash. Integration does not mean an absence of losses: in the MiMo table, some results fall below the specialized teacher. Evaluating retention by domain is more informative than announcing a student that is better at everything. MOPD paper [8].

5. Inference: MTP and cache solve different bottlenecks

In V2-Flash, MTP modules are reused as a drafter for speculative decoding: they propose tokens whose acceptance is checked by the main model. The report publishes decoding acceleration of up to 2.6 times in the evaluated configuration. This number is not a promise for every batch, hardware setup, or input length. V2-Flash report [4].

HySparse, in turn, investigates attention and KV storage. A full-attention layer provides token selection and cache to following sparse layers. The proposal therefore seeks to reduce computation and memory, rather than merely saving operations while keeping independent caches.

The paper evaluates 7B dense and 80B MoE models; for the latter, it reports a reduction in KV close to tenfold. These are experiments for that architecture. HySparse should not be treated as a synonym for V2-Flash's hybrid SWA/global attention, or as demonstrated memory savings for every MiMo. HySparse paper [9].

What lies behind a thousand tokens per second

The V2.5-Pro UltraSpeed announcement combines FP4 quantization of experts with quantization-aware training, a DFlash drafter, and kernels adapted by TileRT. The other modules preserve the original precision. Xiaomi reports more than one thousand decoding tokens per second on a node with eight GPUs.

There are three reading cautions. First, decoding alone does not describe prefill, queue wait, or total latency. Second, a node with eight GPUs is not equivalent to a personal machine. Third, acceptance of proposed tokens varies by task; the article itself acknowledges difficulties in more divergent conversation.

The announced access period was June 9 to 23, 2026. That historical announcement does not prove current commercial availability. The technical interest lies in combining weight representation, speculative proposals, and kernel execution. UltraSpeed technical announcement [10].

6. Agentic RL also depends on who waits for resources

A rollout may depend on code execution on a CPU, evaluation on a GPU, and external services. Reserving everything for the entire trajectory keeps resources occupied even when that stage is not using them.

ARL-Tangram proposes action-level orchestration, with sharing and elasticity for external resources. The central metric includes the time to complete actions, not only utilization of the training GPU. The authors report up to 4.3 times improvement in that measure and up to 1.5 times acceleration of training steps in their scenarios.

A gain for one action does not automatically become the same gain for the complete training run. That mismatch is precisely a reason to measure the whole pipeline. ARL-Tangram paper [11].

7. Tool repetition: measure the path to the answer

Xiaomi documented repetition in V2.6. For a reproducible measure, it counted identical calls, with canonicalized JSON arguments, within one turn before tool feedback. The rate is (N − U) / N: total calls minus unique calls, divided by total.

This measure does not capture loops between turns, near-duplicates, or hidden subcalls in scripts. It also does not confuse repetition with justified retries.

Applying a volume penalty was an attempt with limited generalization. The final correction trained a specialized RL teacher and incorporated the behavior through MOPD. Getting the task right and avoiding redundant actions needed to enter the training process.

The company reports reduced repetition; that does not mean universal absence of loops. Corrected checkpoints use the MOPD suffix, an important distinction when reproducing an evaluation. Diagnosis and correction [12], Pro-MOPD [13], Flash-MOPD [14].

8. Agent memory is state architecture

In MiMo Code, continuity involves a separate writer, structured checkpoints, and context reconstruction. Session, project, and user memories have distinct life cycles; the history allows returning to original records. The design also includes candidate selection, completion verification, and workflows executed in code.

This makes it possible to chain limited physical windows into a longer logical session. It does not turn the model window into infinite memory. Consolidating records and creating skills change the context and procedures available to the agent; they do not, by themselves, update the weights.

This is the most useful technical point for people building agents: continuity requires deciding what to persist, who may write, and how to recover state. A summary may help, but it needs to be part of an operational contract. MiMo Code architecture [15].

9. An open family that goes beyond the main model

The set of papers and projects shows complementary fronts:

MiMo-VL — What it adds to the family: Integrates vision and language; provides SFT and RL checkpoints for the 7B model.; Official material: MiMo-VL repository [16]

MiMo-Audio — What it adds to the family: Investigates audio understanding and generation, including few-shot learning.; Official material: MiMo-Audio repository [17]

MiMo-V2.5-ASR — What it adds to the family: Speech recognition with documented focus on Mandarin, English, Chinese dialects, and difficult acoustic conditions.; Official material: MiMo-V2.5-ASR repository [18]

MiMo-V2.6 — What it adds to the family: Pro and Flash omnimodal models, with public weights and documentation.; Official material: Announcement and materials [1]

This path can be followed through the papers [19] and technical blog [20] sections. The contribution lies both in the models and in the possibility of examining engineering choices and problems. Releasing a correction and explaining the failure that motivated it also produces useful knowledge for people building other systems.

The Distill-Qwen-9B example

MiMo-V2.6-Distill-Qwen-9B makes this circulation of knowledge especially concrete. According to the official card, it was developed by SFT of Qwen3.5-9B with data generated by MiMo, as an open starting point for research in agentic RL. The card declares an MIT license.

It is a 9B checkpoint with an explicit technical origin. This allows investigating transfer of capabilities to another scale without confusing it with an automatically reduced version of the 1.02T MoE. It also does not establish that other MiMo models have the same Qwen base. Model and training description [21].

10. Voice: a front we want to hear in Portuguese

The MiMo-V2.5-TTS family includes synthesis with ready-made voices, voice creation from a description, and cloning from samples. Expressive control through instructions and audio markers allows work on rhythm, emotion, and intonation. It is an interesting front for agents whose interaction needs to go beyond text. TTS presentation [22].

Here is one of our wishes: to try this proposal in Brazilian Portuguese. In the documentation consulted, ready-made voices are identified in Chinese and English; we did not find officially documented Portuguese support. We hope to see our language included, especially because an expressive voice could add to accessibility. We have not yet performed our own evaluation of audio quality. Synthesis and voice documentation [23].

The TTS is presented in these materials as an API service. Its availability should not be confused with the release of weights for other models: openness must be identified for each component.

What we need to measure in a real system

Our reading is that an agent evaluation should bring together outcome, trajectory, and continuity. In addition to task tests, it is useful to record redundant calls, context consumption, time without progress, cost per completed task, and fidelity to constraints after context reconstruction.

To reproduce a comparison, the record must identify checkpoint and revision, tokenizer and template, precision, inference engine, hardware, generation parameters, harness, and task set. It must separate evaluation of an isolated run from tasks with multiple interactions, including failures and result dispersion.

The V2.6 cards consulted declare MIT. This does not authorize assigning a single license to the whole ecosystem: the V2-Flash GitHub repository, for example, presents Apache-2.0. The cited artifact must be identified. Pro-MOPD card [13], V2-Flash repository [24].

At Kuaapy, this investigation connects to the experience of people who depend on an agent to complete a task: how many decisions need to be repeated, how many corrections are necessary, and how much human effort remains after supposed automation. These are engineering and accessibility questions that can guide experiments without anticipating their results.

The openness of the models is therefore a central part of the value we recognize in Xiaomi MiMo. Having weights, documentation, and references available expands the space for examining and developing technology, including outside large laboratories. This is also a direction we consider for models we may develop at Kuaapy. The formats and conditions of any eventual release have not yet been defined.

Learn about Xiaomi MiMo from official sources

To follow the team's work, consult the models, and explore the original materials:

• Official Xiaomi MiMo site [25] — news, research, and products.

• Xiaomi MiMo on Hugging Face [2] — models, technical cards, and collections published by the team.

• Xiaomi MiMo on GitHub [26] — official project repositories.

• MiMo Code [27] — official programming-agent page.

• Team papers [19] and technical blog [20] — indexes for following the research.

• MiMo-V2.5-TTS family [22] and speech-synthesis documentation [23].

Papers and technical documentation for this article

• MiMo-7B: from pretraining to post-training [28].

• MiMo-V2-Flash — technical report [4].

• Rollout Routing Replay — router alignment [29].

• MOPD — on-policy distillation with multiple teachers [30].

• HySparse — sparse attention and KV sharing [9].

• ARL-Tangram — infrastructure for agentic RL [11].

• MiMo Code — architecture for long-horizon tasks [15].

• Diagnosis and correction of repeated calls [12].

• UltraSpeed — TileRT codesign [10].

Cited models and corrected checkpoints

• MiMo-V2.6-Pro-RL [5].

• MiMo-V2.6-Flash-RL [6].

• MiMo-V2.6-Pro-MOPD [13].

• MiMo-V2.6-Flash-MOPD [14].

• Official MiMo-V2-Flash repository [24].

• MiMo-V2.6-Distill-Qwen-9B [21].

• MiMo-VL [16], MiMo-Audio [17], and MiMo-V2.5-ASR [18].