Xiaomi's MiMo Gets Its Government Stamp — and a Flagship Phone to Match

Xiaomi's MiMo Gets Its Government Stamp — and a Flagship Phone to Match

On July 15, 2026, China's Cyberspace Administration published a list that quietly rewrote the rules of the country's AI hardware race. Seven on-device generative AI services had cleared the national registration process required under the *Interim Measures for the Administration of Generative Artificial Intelligence Services* — and sitting alongside Apple, Huawei, OPPO, vivo, Samsung, and Nubia was Xiaomi, with its in-house MiMo AI framework officially registered as "Xiaomi Surge AI."

For a company that only entered the large-model race in earnest in late 2025, the milestone is striking. Xiaomi now holds the same regulatory clearance as the country's most established AI incumbents, and it is moving fast to make that clearance count: a trillion-parameter open-weight model, a terminal-native coding assistant, and a foldable flagship phone designed to run the whole stack are all either live or imminent. The MiMo story is no longer a footnote in China's AI narrative — it is becoming a chapter of its own.

---

The Registration That Changes the Game

The July 15 batch approval was the result of filings submitted on July 8, 2026, and it matters for a specific reason: without CAC registration, a generative AI service cannot be offered to the Chinese public at scale. The approval is not a rubber stamp — it requires demonstrating compliance with content safety rules, data localisation requirements, and algorithmic transparency obligations.

The seven approved services span the full spectrum of China's smartphone ecosystem:

• Apple Intelligence (registered via Apple Technology Development Shanghai) — powered by Alibaba's Qwen for language and Baidu for visual search, as TechTimes reported↗ after a 22-month approval process
• Huawei Xiaoyi AI Large Model — running on Ascend silicon, deeply integrated with HarmonyOS
• Xiaomi Surge AI / MiMo — Xiaomi's full model matrix, covering language, multimodal, and voice
• OPPO AndesGPT Large Model — OPPO's proprietary on-device system
• vivo Blue Heart On-Device Large Model — vivo's edge inference stack
• Samsung Galaxy AI (registered via Samsung China Investment) — the localised version of Galaxy AI
• Nubia Doubao Phone Large Model — ByteDance's Doubao running on Nubia hardware

"This concentrated batch of approvals is a watershed moment that clears the path for a new wave of AI-powered smartphone launches expected to intensify in the second half of 2026." — Industry analysts cited by Pandaily↗

For Xiaomi specifically, the registration unlocks the ability to deploy MiMo natively across its entire device ecosystem — smartphones, smart home hardware, and eventually its electric vehicles — without the legal uncertainty that has constrained other AI features in China.

---

What MiMo-V2.5-Pro Actually Is

The model at the centre of Xiaomi's AI ambitions is MiMo-V2.5-Pro, released on April 22, 2026, and open-sourced under the MIT License on April 27. The official model page↗ describes a 1.02-trillion-parameter Mixture-of-Experts architecture with 42 billion active parameters per inference — a design that keeps compute costs manageable while delivering frontier-class reasoning depth.

Technical Specifications at a Glance

• Total parameters: 1.02 trillion (MoE), 42B active per token
• Context window: 1,048,576 tokens (approximately 1 million)
• Architecture: Hybrid attention — 6:1 ratio of Sliding Window Attention to Global Attention, reducing KV-cache storage by roughly 7× compared to full-attention designs
• Training data: ~27 trillion tokens, FP8 mixed precision
• Multimodal support: Text, image, audio, and video natively
• License: MIT (full commercial use, fine-tuning, secondary training permitted)
• Weights: Available on Hugging Face↗ with Day-0 support for vLLM and SGLang

The companion base model, MiMo-V2.5↗, is a 310-billion-parameter native omnimodal model with 15B active parameters — lighter, faster, and designed for the kinds of real-time tasks that run on-device.

Benchmark Performance

Xiaomi has been careful not to over-claim on benchmarks, but independent evaluations have been broadly positive. On the BenchLM public leaderboard↗, MiMo-V2.5-Pro holds an overall score of 69.38, ranking #17 out of 215 models as of late July 2026. On Artificial Analysis↗, it sits at #6 out of 97 on the Intelligence Index.

For coding specifically — the use case Xiaomi has leaned into hardest — the numbers are competitive:

• SWE-bench Pro: 57.2% (versus GPT-5.5 at 58.6% and Claude Opus 4.8 at approximately 60%)
• SWE-bench Verified: 78.9%
• Terminal-Bench 2.0: 68.4
• Claw-Eval: 64% Pass³ rate

Perhaps more telling than the benchmark scores are the real-world demonstrations Xiaomi has published: the model autonomously implemented a complete SysY compiler in Rust in 4.3 hours across 672 tool calls with a 100% test pass rate, and built a multi-track video editor in 11.5 hours across 1,868 tool calls. These are not cherry-picked demos — they reflect a model genuinely capable of sustaining coherent, long-horizon agentic workflows.

The key differentiator is what Xiaomi calls "harness awareness" — the model's ability to actively manage its own context and tool usage across thousands of sequential steps, rather than losing coherence as the context window fills.

---

MiMo Code: The Developer-Facing Layer

The model alone is not the product. On June 10, 2026, Xiaomi released MiMo Code, an open-source terminal-native coding assistant built on the OpenCode framework and licensed under MIT. The official release notes↗ describe it as a direct competitor to Claude Code, with a persistent memory architecture that addresses the "AI amnesia" problem that plagues most coding agents.

What Makes MiMo Code Different

• Persistent memory via SQLite FTS5: Four-layer memory system covering project context, session checkpoints, scratch notes, and task progress logs — managed by an independent "checkpoint-writer" subagent so the primary agent never has to context-switch
• Compose Mode: Tab-activated autonomous development mode where the agent handles design, planning, coding, testing, and review end-to-end
• Self-evolution via `/dream`: Every seven days, the tool automatically deduplicates and compresses historical sessions into long-term memory
• Voice control: Powered by MiMo-V2.5-ASR with TenVAD, enabling hands-free operation including verbal command correction
• Multi-model support: Works with DeepSeek, Kimi, and GLM backends in addition to MiMo's own models

In internal benchmarks, MiMo Code achieved a 62% success rate on SWE-Bench Pro (versus 57% for Claude Code) and 73% on Terminal Bench 2 (versus 68% for Claude Code) when using the same underlying model. The tool is installable in a single command on macOS and Linux (`curl -fsSL mimo.xiaomi.com↗ | bash`) and via npm on Windows.

As of late July 2026, Xiaomi has transitioned MiMo Code's free-access tier to a subscription-based Token Plan — a sign that the product has moved from developer acquisition mode into commercial operation.

---

Pricing and Access: Where MiMo Sits in the Market

For developers who want API access without self-hosting, MiMo-V2.5-Pro is available on OpenRouter↗ at $0.43 per million input tokens and $0.87 per million output tokens. Direct access through Xiaomi's own platform can be cheaper — some providers have listed input prices as low as $0.348/M tokens — and the market has seen significant price deflation since launch, with the cheapest available input price dropping approximately 65% over the 90 days following release.

For context, that puts MiMo-V2.5-Pro at roughly:

• 3–4× cheaper than Claude Opus 4.8 on a per-token basis
• Comparable to DeepSeek-V4-Pro for standard workloads, with higher cache-miss costs but lower effective costs for novel context
• Significantly cheaper than GPT-5.5 for long-horizon agentic tasks where token efficiency matters

The MIT license means that teams with sufficient GPU infrastructure can self-host the full model and eliminate per-token costs entirely — though the 1T-parameter scale requires serious multi-GPU setups. The MiMo-V2.5-Pro-FP4-DFlash↗ quantized variant, which uses MXFP4 quantization and block-diffusion speculative decoding, brings the memory footprint down substantially while maintaining near-lossless quality.

---

The Hardware Endgame: Xiaomi 18 Fold

The regulatory clearance and the model stack are both building toward a single product moment: the Xiaomi 18 Fold, expected to launch in August 2026. Regulatory filings for the device (model number 2608BPX34C, codenamed LHASA) reveal an AI integration strategy that goes well beyond a single model.

According to XimiTime's analysis of the filings↗, the 18 Fold will support a multi-model architecture that includes:

• Xiaomi's own MiMo Large Model, HyperVision Large Model, and HyperPerception Large Model
• Third-party integrations with Zhipu AI, Doubao (ByteDance), DeepSeek, Tongyi Qianwen (Alibaba), and Hunyuan (Tencent)

This is not a single-model phone — it is a model router in a foldable chassis, delegating tasks to whichever model is best suited for the job. The device will run on Xiaomi's in-house XRING O3 chipset (TSMC 3nm), which is specifically designed for the "triple integration" of proprietary silicon, HyperOS 4, and MiMo AI models.

The 18 Fold is expected to be priced around CNY 10,000 (~$1,480) and will initially be China-only — but it represents the clearest statement yet of where Xiaomi sees the AI phone category heading: not a phone with a chatbot bolted on, but a device where the AI layer is as fundamental as the operating system.

---

Why This Matters Beyond China

Xiaomi's MiMo story is easy to underestimate if you're watching from outside China. The company is not a frontier AI lab in the DeepSeek or Moonshot sense — it does not publish research papers at the same cadence, and its models have not generated the same global headlines. But the combination of factors at play here is genuinely significant for the global developer community.

First, the MIT license on MiMo-V2.5 and MiMo-V2.5-Pro means these are among the most permissively licensed trillion-parameter models available anywhere. Developers can fine-tune, redistribute, and build commercial products on top of them without restriction — a meaningful contrast to the modified licenses that govern Kimi K3 and some Qwen variants.

Second, the MiMo Code toolchain is a serious attempt to build a vertically integrated developer experience: model, coding assistant, and API platform all from one vendor, all MIT-licensed, all interoperable. That is a proposition that competes directly with Anthropic's Claude Code ecosystem and GitHub Copilot, not just with other Chinese models.

Third, Xiaomi's $23 billion USD AI investment commitment for 2026 alone — backed by a hardware distribution network that ships hundreds of millions of devices annually — gives MiMo a deployment surface that pure-play AI labs cannot match. When MiMo clears regulatory hurdles in China, it does not just unlock API access; it unlocks native integration across an installed base of smartphones, smart TVs, and connected home devices.

For global developers evaluating Chinese open-weight models, MiMo-V2.5-Pro deserves a place on the shortlist alongside DeepSeek-V4-Pro and GLM-5.2 — not because it tops every benchmark, but because the full ecosystem around it is unusually coherent for a model that is less than four months old.

The national registration milestone is, in one sense, a bureaucratic checkbox. In another, it is the moment Xiaomi's AI ambitions became legally real — and the 18 Fold, MiMo Code, and a trillion-parameter open-weight model are the evidence that the company intends to make the most of it.

---

How to Try It

• API access: OpenRouter↗ ($0.43/$0.87 per M tokens input/output) or Xiaomi's native Token Plan
• Self-hosting: Weights on Hugging Face↗ under MIT license; FP4 quantized variant available for reduced VRAM requirements
• MiMo Code: Install via `curl -fsSL mimo.xiaomi.com↗ | bash` (macOS/Linux) or `npm install -g @mimo-ai/cli` (Windows); MIT licensed, multi-model backend support
• Benchmarks and comparisons: BenchLM↗ and Artificial Analysis↗ both maintain live rankings