Home / Blog / MiMo On-Device Compliance

Published: July 16, 2026 · Last reviewed: July 20, 2026

MiMo On-Device Compliance: China's First 7-Phone Edge-LLM Filing

What happened

China's Interim Measures for the Management of Generative AI Services (effective August 2023) require that any generative AI service with public-opinion or social-mobilization characteristics go through an algorithm filing before commercial launch. The first phone-side (on-device) filings under this regime were published on July 15, 2026. The full list:

Brand	Model name	Company	Region	Notes
Xiaomi	小米澎湃AI (MiMo)	小米科技有限责任公司	Beijing	Filing no. Beijing-XiaomiPengPaiAI-202509080119
Apple	Apple 智能 (Apple Intelligence)	Apple 技术开发(上海)有限公司	Shanghai	On-device LLM partner: Alibaba Qwen
Huawei	华为小艺 AI 大模型	华为软件技术有限公司	—	Fully self-developed
OPPO	AndesGPT	OPPO 广东移动通信有限公司	Guangdong	Fully self-developed
vivo	vivo 蓝心端侧大模型	维沃移动通信有限公司	Guangdong	Fully self-developed
Samsung	Galaxy AI (盖乐世 AI)	三星(中国)投资有限公司	—	First filing of Samsung's CN-side model
ZTE / Nubia	努比亚豆包大模型	上海申启纪元智能终端有限责任公司	Shanghai / Jiangsu	On-device LLM: ByteDance Doubao

All 7 were filed on 2026-07-08. The CAC announcement explicitly said this is the first batch of phone-side generative AI filings — prior filings had been cloud-side (server-deployed) models. On-device deployment is a different compliance path because the model runs entirely on user hardware, so the regulatory question is more about output safety, content filtering, and update channels than about server-side data handling.

What Xiaomi's filing covers

Per Xiaomi's own statement (published through the official MiMo tech account on July 16):

• Full MiMo model matrix closed-loop: LLM + multimodal + speech. All three layers are MIT-licensed open weights on HuggingFace, and the on-device production deployment uses the same family.
• End-cloud collaboration as the deployment pattern: lightweight inference on-device, heavier reasoning offloaded to Xiaomi's cloud cluster. This is the same architecture as Apple Intelligence's "Private Cloud Compute" tier.
• Human–Car–Home coverage: the same MiMo family ships in phones (HyperOS), cars (YUU7 cockpit, SU7 cockpit), and smart home devices. The on-device filing is the first regulatory checkpoint in that whole rollout.
• 2026 R&D: ¥16B. 3-year total: ≥¥60B in AI across chips, OS, and model layers.

Lei Jun's framing: "AI, chips, and OS are Xiaomi's three core long-term bets." The 7-8 filing is the first tangible output of that bet on the model side.

Why the filing matters beyond compliance

For developers: MiMo is now safe to integrate into commercial CN-side products without a regulatory gray zone. App-store distribution, pre-installation deals, B2B SDK sales — all clear to ship without waiting for a separate filing. That's a real unlock for anyone building on top of MiMo in CN.

For the on-device LLM industry: the first batch validates a regulatory model — cloud-side filings since 2023, phone-side now in 2026. Other OEMs in the second tier (Honor, Transsion, Lenovo) are likely to file in the next 6–12 months. The "who gets to be on-device LLM" race has moved from product to compliance.

For MiMo's model strategy: the LLM + multimodal + speech matrix is what made the filing possible. A model that only does text would have needed separate multimodal and speech filings. The closed-loop claim is both a product advantage and a regulatory efficiency.

Open questions

• What about the cloud-side MiMo API? Xiaomi runs MiMo-V2.5 standard and V2.5-Pro on platform.xiaomimimo.com (per our pricing page). The cloud-side model has been in production since April 2026; whether it has a separate CAC filing or operates under the same parent filing isn't publicly disclosed.
• What about MiMo-7B self-hosters? MIT-licensed open weights fall outside CAC's scope — the regulator covers service providers, not end users who self-host a model on their own hardware. The filing matters for anyone shipping a product with MiMo embedded.
• What about models that ship pre-installed via the AOSP stack? Honor and Lenovo are missing from the list but have shipped AI features. Likely next batch.
• When does the on-device AI feature actually launch to users? Per a CAC source quoted in coverage, "filing is not equivalent to feature launch." Expect the first end-user-facing features to roll out via OTA in late July to early August 2026.

What's in it for non-CN developers

The CN-side filing doesn't affect international MiMo use. MiMo remains MIT-licensed, weights are on HuggingFace, the API is OpenAI-compatible, and the international pricing (¥3 / ¥6 per 1M for V2.5-Pro list) is the same. The 7-8 filing is a CN-specific regulatory milestone, not a model capability change.

What it does signal internationally: Chinese OEMs are committing to a fully in-house AI stack. The 7 filings include 4 fully self-developed models (Xiaomi, Huawei, OPPO, vivo) and 3 partner-assisted (Apple+Qwen, Samsung+Galaxy, Nubia+Doubao). The split suggests domestic OEMs are more willing to invest in their own model training pipelines, while international brands partner with local AI vendors for CN-side deployment.

Related pages

• On-device LLM deployment guide — running MiMo-7B on phones, cars, IoT
• MiMo-V2.5 Series deep dive — the LLM + multimodal + speech family
• MiMo ecosystem research — the Human–Car–Home rollout
• Xiaomi-Robotics-U0 (July 15, 2026) — the embodied foundation model that dropped the same week

Unofficial community write-up of the July 15–16, 2026 CAC publication and Xiaomi's official statement. For the canonical filing data, see the CAC public registry.