📱Ifanr (爱范儿)•Stalecollected in 26m
Xiaomi: The Token Price Butcher
💡Xiaomi's aggressive price cuts are shaking up the AI infrastructure market. Check if you can save on inference costs.
⚡ 30-Second TL;DR
What Changed
Aggressive reduction in token pricing for AI services
Why It Matters
Forces other AI providers to re-evaluate their pricing models to remain competitive in the mass-market developer space.
What To Do Next
Compare Xiaomi's new token pricing against your current LLM provider to see if switching can reduce your infrastructure overhead.
Who should care:Founders & Product Leaders
Key Points
- •Aggressive reduction in token pricing for AI services
- •Strategy focused on high-volume, low-cost AI accessibility
- •Direct competition with existing cloud and model providers on price
🧠 Deep Insight
Web-grounded analysis with 22 cited sources.
🔑 Enhanced Key Takeaways
- •Xiaomi has open-sourced its MiMo-V2.5 and MiMo-V2.5-Pro models under the permissive MIT License, making them suitable for commercial applications and positioning them as foundational infrastructure for the next generation of AI agents.
- •The MiMo-V2.5-Pro model, with 1.02 trillion parameters, demonstrates high token efficiency, requiring approximately 40-60% fewer tokens than comparable frontier models like Anthropic Claude Opus 4.6, Google Gemini 3.1 Pro, and OpenAI GPT-5.4 for agentic tasks.
- •Xiaomi's AI strategy is deeply integrated into its 'Human x Car x Home' smart ecosystem, aiming to extend AI beyond screens into physical world interactions through products like Xiaomi Miloco, which enables intelligent automation across smart home devices and electric vehicles.
- •The company has committed a substantial investment of over US$8.8 billion (60 billion yuan) in AI over the next three years, underscoring its ambition to develop frontier AI models and capabilities.
- •Xiaomi's pricing strategy includes not only aggressive API token cost reductions (up to 99% for cache hits) but also an overhauled 'Token Plan' subscription offering 5-8 times more usable tokens for the same price, and unified pricing across all context lengths.
📊 Competitor Analysis▸ Show
| Model / Provider | Input (Cache Hit) / Million Tokens | Input (Cache Miss) / Million Tokens | Output / Million Tokens | Key Features & Benchmarks |
|---|---|---|---|---|
| Xiaomi MiMo-V2.5-Pro | $0.0036 (RMB0.025) | $0.435 (RMB3) | $0.87 (RMB6) | 1.02T parameters, 1M context window, 40-60% more token efficient for agentic tasks than competitors. |
| Xiaomi MiMo-V2.5 | $0.0028 (RMB0.02) | $0.14 (RMB1) | $0.28 (RMB2) | 310B parameters, 1M context window, native omnimodal capabilities. |
| DeepSeek V4 Pro | $0.0036 (RMB0.025) | $0.435 (RMB3) | $0.87 (RMB6) | Xiaomi aligned pricing directly with DeepSeek V4 Pro. |
| DeepSeek V4 Flash | $0.0028 (RMB0.02) | $0.14 (RMB1) | $0.28 (RMB2) | Xiaomi aligned pricing directly with DeepSeek V4 Flash. |
| ByteDance Doubao (Seed-2.0-Pro) | N/A | $0.46 (RMB3.2) | $2.32 (RMB16) | Part of the 'price-cut camp' leveraging large ecosystems. |
| Kimi Moonshot V1 | N/A | $1.45 (RMB10) | $4.35 (RMB30) | Higher pricing, justified by long context and inference capabilities, achieved counter-trend rises in usage. |
| Anthropic Claude Opus 4.6 | N/A | N/A | $25.00 | Mentioned as requiring 40-60% more tokens for comparable agentic tasks than MiMo-V2.5-Pro. |
| Google Gemini 3.1 Pro | N/A | N/A | N/A | Mentioned as requiring 40-60% more tokens for comparable agentic tasks than MiMo-V2.5-Pro. |
| OpenAI GPT-5.4 | N/A | N/A | N/A | Mentioned as requiring 40-60% more tokens for comparable agentic tasks than MiMo-V2.5-Pro. |
🛠️ Technical Deep Dive
- MiMo-V2.5-Pro: Features 1.02 trillion total parameters with 42 billion active parameters, utilizing a Mixture-of-Experts (MoE) architecture. It supports a 1-million-token context window and employs a Hybrid Attention mechanism with a 7:1 ratio of Sliding Window Attention (SWA) to Global Attention (GA) for efficiency. A Multi-Token Prediction (MTP) module is integrated for faster generation, optimizing it for complex software engineering and long-horizon agentic tasks.
- MiMo-V2.5: A 310-billion-parameter model with native omnimodal capabilities, processing text, images, video, and audio within a single architecture. It also features a 1-million-token context window and delivers agentic performance comparable to its Pro sibling at approximately half the token cost.
- MiMo-V2-Flash: An open-weight Mixture-of-Experts (MoE) model with 309 billion total parameters and 15 billion active parameters. Its architecture includes a hybrid attention strategy (5:1 SWA to GA ratio) to reduce KV cache storage requirements by nearly 6x. It incorporates a learnable attention sink bias for coherence over sequences up to 256,000 tokens and uses Rollout Routing Replay (R3) to ensure training-inference consistency. A Multi-Token Prediction (MTP) module is embedded for self-speculative decoding, effectively tripling inference speeds.
🔮 Future ImplicationsAI analysis grounded in cited sources
Xiaomi's aggressive pricing and open-source strategy will intensify the global AI price war, particularly for agentic models.
By offering high-performance, token-efficient models at significantly lower costs and under permissive licenses, Xiaomi forces competitors to re-evaluate their own pricing and accessibility strategies to retain market share.
The integration of AI across Xiaomi's 'Human x Car x Home' ecosystem will accelerate the adoption of pervasive AI in consumer devices and smart environments.
Xiaomi's strategy to embed its MiMo models into smartphones, EVs, and smart home devices, coupled with user-friendly interfaces like Xiaomi Miloco, will make advanced AI capabilities a seamless part of daily life.
Xiaomi will emerge as a dominant player in the open-source AI community, especially for agent-focused applications.
Releasing powerful models like MiMo-V2.5-Pro under the MIT license, combined with high token efficiency and competitive benchmarks for agentic tasks, positions Xiaomi as a key enabler for developers building the next generation of AI agents.
⏳ Timeline
2010-04
Xiaomi founded, later launching its MIUI Android-based firmware.
2023-10
Xiaomi upgraded its corporate strategy to the 'Human × Car × Home' smart ecosystem.
2025-04
Xiaomi announced its venture into large-scale AI models with the introduction of Xiaomi MiMo.
2026-03
Xiaomi committed to investing at least RMB60 billion (US$8.8 billion) in AI over the next three years.
2026-04-22
Xiaomi released MiMo-V2.5 and MiMo-V2.5-Pro, with the Pro version being open-sourced under the MIT license.
2026-05-27
Xiaomi announced permanent API price reductions for its MiMo-V2.5 series, cutting costs by up to 99% for cache hits and unifying pricing across context lengths.
📎 Sources (22)
Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.
📰
Weekly AI Recap
Read this week's curated digest of top AI events →
👉Related Updates
AI-curated news aggregator. All content rights belong to original publishers.
Original source: Ifanr (爱范儿) ↗