On May 27, Xiaomi announced permanent price cuts for the MiMo-V2.5 series API, with reductions of up to 99%, along with 5-8x larger Token Plan allowances. It is the next large-model vendor after DeepSeek to go with "permanent price cuts," and the two domestic models now price almost head-to-head.

Price Cut Details

The new pricing covers both MiMo-V2.5 and MiMo-V2.5-Pro, and took effect globally at 0:00 Beijing time on May 27. The most important change: context window length is no longer differentiated — 256K and 1M now share one price.

API Price Comparison

💡 Tip: The 99% cut mostly shows up in "cache hit" scenarios. Highly repetitive contexts, high-frequency Agents, multi-round code tasks, and batch inference tasks benefit most from the low prices. Applications with low cache-hit rates will never touch the price floor.

Token Plan Changes

Pricing unchanged, credits greatly increased:

Calculated for scenarios with 95%+ cache-hit rates, the 39-yuan Lite plan can run more than 500 million tokens on MiMo-V2.5, and the 659-yuan Max plan breaks 10 billion tokens. Even on the pricier MiMo-V2.5-Pro, the Lite plan still reaches more than 190 million tokens.

For all Token Plan users with subscriptions still active, credits were reset at 0:00 on May 27 and automatically switched to the new billing rules.

Why It Can Cut Prices

Xiaomi's announcement explained the inference optimizations behind it:

• Full SWA (sliding window attention) support based on SGLang HiCache, cutting the amount of KV Cache data moved between GPU VRAM, CPU memory, and SSD down to about 1/7 of pre-optimization levels
• The number of cacheable tokens increased to roughly 5x
• Optimized expert parallelism and input-length bucketing strategies to boost cluster input throughput

Behind the price war is, at its core, a competition in inference systems and infrastructure capability.

Where the MiMo-V2.5 Series Models Sit

MiMo-V2.5-Pro ranks tied for first among global open-source models on the Artificial Analysis overall intelligence leaderboard, and ties for first among global open-source models on the Agent index as well. It is open-sourced under the MIT license and supports commercial deployment and secondary training.

What It Means for Developers

1. Agent and coding assistant costs plummet: Cache-hit rates are high in Agent scenarios, so MiMo's low pricing is a natural fit for high-frequency multi-round tasks
2. Roughly on par with DeepSeek pricing: The two domestic models have formed a price anchor, and developers can switch flexibly based on model capability
3. Overseas users benefit too: Prices are synced globally, and the overseas community response has been positive — "MiMo scores 3 points below Opus at more than a hundred times cheaper"

How to Get It

• API calls: Visit Xiaomi's open platform to get an API Key and call directly at the new prices
• Token Plan: Starting at 39 yuan, suited to developers with steady usage needs
• Open-source deployment: The MiMo-V2.5 series is open-sourced under the MIT license and can be self-deployed