Xiaomi

A 9B agentic model Xiaomi ships as a shared starting point for open research on agentic RL, alongside the MiMo-V2.6 weights, roughly 7,000 released RL…

The efficiency-balanced sibling of MiMo-V2.6-Pro, released the same day as open weights under MIT. 309B total / 15B active sparse MoE: 48 layers (39 s…

Xiaomi's flagship omnimodal reasoning model, released as open weights under MIT on September 21, 2026 after a publicly livestreamed RL run. It debuted…

Xiaomi's tabular foundation model, an in-context learner for classification and regression pretrained exclusively on synthetic data from structural ca…

Xiaomi's omnimodal sparse-MoE model (310B total / 15B active), distinct from the MiMo-V2.5-Pro flagship: hybrid attention (sliding-window + global), 1…

Open-source terminal AI coding agent with cross-session memory — Xiaomi's entry into the coding-agent CLI space, built as a fork of OpenCode (MIT, Typ…

1.02T total / 42B active MoE with hybrid attention (sliding window + global at 6:1 ratio, 128-token window) and Multi-Token Prediction. 1M-token conte…

Xiaomi's first fully omni-modal foundation model. Native multimodal architecture that can "see, hear, and act," designed to power Xiaomi's next-genera…

Xiaomi's 1-trillion parameter MoE flagship (42B active) designed for agentic workloads with a 1-million-token context window. Before its official reve…

High-fidelity text-to-speech model for providing agent voices across Xiaomi's device ecosystem.

4.7B parameter Vision-Language-Action (VLA) model for real-time robotic execution. Scales VLA models for consumer GPUs.

309B MoE model (15B active) optimized for extreme inference speed (150 tokens/sec) and low-cost API access. Introduces a hybrid attention architecture…

Library/technique for distilling intelligence from multiple teacher models into smaller student models.

First cross-embodied foundation model integrating autonomous driving and embodied AI. SOTA on 17 embodied AI and 12 autonomous driving benchmarks.

Framework enabling LLM-powered smart home automation via local natural language interaction.

"Stabilizing MoE Reinforcement Learning by Aligning Training and Inference Routers" — from LLM-Core Xiaomi (with Peking University; senior autho…

First operating system with a system-level AI layer (HyperAI) integrated across smartphones, Xiaomi SU7/SU8 vehicles, and IoT devices.

38,000-hour collection of general audio captions for training holistic audio-reasoning models.

Open-source vision-language model that outperformed models ten times its size (like Qwen-72B) in multimodal reasoning. Includes SFT and RL variants.

Specialized model for few-shot audio understanding and environmental sound classification. Pre-trained on 100M+ hours, achieves SOTA on speech intelli…

3nm custom AI chip designed for on-device reasoning in Xiaomi 15/16/17 series.

Reasoning-focused model achieving superior performance on AIME (math) and HumanEval (coding). From pretraining to posttraining, unlocking reasoning po…

Curated sets of mobile application screenshots and UI hierarchies for VLM training.

Vision-language model specialized in understanding mobile UI/UX layouts.

Audio encoder and language model family. Original Dasheng encoder (1.2B params, 272K hours) evolved into DashengLM and MiDashengLM-7B for efficient au…

Xiaomi's first significant LLM, ranked top in C-Eval and CMMLU at release.

Outlined Xiaomi's ethical AI principles: Security, Privacy, Fairness.

Xiaomi's most famous open-source project. Deep learning inference framework optimized for heterogeneous computing (CPU/GPU/DSP) on mobile devices.

Foundational voice assistant that served as the initial testbed for Xiaomi's NLP and speech research.