Xiaomi and MiniMax Launch Major Moves Simultaneously, Agent Pricing War Officially Begins

On March 18 and 19, two Chinese companies each released their own large models focused on Agent technology. Domestic AI startup MiniMax launched M2.7, and Xiaomi’s large model team MiMo released V2-Pro. Both models ranked in the top tier globally on the Agent benchmark, but their API output pricing is $1/21 and $1/8 of Claude Opus 4.6, respectively.

Both companies made their moves in the same week, but their strategies are completely different. They represent two distinct technological paths, betting on two futures for the Agent era.

The Same Exam, 1/17 Tuition

Let’s start with the most straightforward comparison.

Based on data from OpenRouter and official pricing pages, the API output prices (per million tokens) are $1.2 for MiniMax M2.7 and $3 for MiMo V2-Pro. For reference, Claude Opus 4.6 costs $25, GPT-5.2 costs $14, and Claude Sonnet 4.6 costs $15.

The price gap is orders of magnitude, but the capability gap is not. On SWE-bench Verified (currently the most mainstream benchmark for measuring coding ability), MiMo V2-Pro scores 78%, Sonnet 4.6 scores 79.6%, less than two percentage points apart. M2.7’s SWE-Pro score is 56.22%, on par with GPT-5.3-Codex. On VIBE-Pro (end-to-end project delivery capability), M2.7 scores 55.6%, close to Opus 4.6.

The focus of this chart isn’t on who’s higher or lower—since benchmark systems aren’t fully aligned, direct comparisons should be cautious. The key point is the “price-performance gap”: domestic Agent models have entered the same capability range but are in completely different price brackets.

Trillion Parameters vs Self-Evolution

Price is just the surface. The two companies have taken two completely different approaches.

MiMo V2-Pro follows a “big effort, big results” strategy. According to Xiaomi’s official announcement, V2-Pro has over 1 trillion total parameters, with 42B active parameters, supporting a context length of 1 million tokens. Its core innovation is the Hybrid Attention mechanism, which combines sliding window attention (SWA) and global attention (GA) at a ratio of 7:1—compared to 5:1 in the previous V2-Flash. This architecture makes the model more stable when handling long documents and multi-tool parallel calls in Agent scenarios. On PinchBench (Agent tool calling capability evaluation), MiMo V2-Pro scores 84%.

M2.7 takes a completely different route. According to MiniMax’s official technical blog published on March 18, the number of parameters for M2.7 has not been disclosed, but it demonstrates a “self-iterative evolution” mechanism: the model autonomously runs over 100 optimization cycles, including analyzing failure trajectories, planning modifications, changing its own code architecture, running evaluations, and looping again, ultimately achieving a 30% performance improvement on internal evaluation sets. In the 22 challenging problems of MLE Bench Lite (machine learning competition difficulty assessment), M2.7 earned 9 gold, 5 silver, and 1 bronze medals, with an average medal rate of 66.6%.

From five dimensions, the two paths’ strengths are completely different: MiMo V2-Pro clearly excels in context length and coding engineering, while M2.7 leads in office automation and self-iteration capabilities. According to the same MiniMax technical blog, M2.7 scored 1495 on GDPval-AA (office document processing evaluation), ranking first among open-source models, and maintained a 97% skill adherence rate in the MM-Claw test covering over 40 complex skills.

Five Months, Four Versions

Not only are their technical routes different, but their iteration rhythms are also entirely distinct.

According to public release records, MiniMax released M2 in October 2025 and M2.7 in March 2026, iterating four major versions within five months, averaging a major release every 49 days. The interval between M2.5 and M2.7 was only about 30 days.

Xiaomi’s MiMo pace is different: in April 2025, it released MiMo-7B (an open-source reasoning model with 7B parameters), followed by V2-Flash (with 309B total parameters) in December, and V2-Pro (1T total parameters) in March 2026. Each generation involves a larger parameter scale, but the intervals are longer.

MiniMax opts for small, rapid iterations, with each update making modest improvements but at a very high frequency. M2.7’s self-iteration mechanism is designed for “continuous evolution.” Xiaomi, on the other hand, chooses a big leap each time, with significant jumps in parameters and architecture.

Anonymous 8 days, tops OpenRouter

Beyond technical routes, Xiaomi’s release strategy also breaks industry norms.

According to Reuters, on March 11, an anonymous model called Hunter Alpha appeared on OpenRouter, the world’s largest API aggregation platform. Without branding, press conferences, or technical blogs, it was priced extremely low but performed surprisingly well.

The community began speculating about its origin. According to Republic World and several tech media outlets, the most mainstream guess is DeepSeek V4, as MiMo team leader Luo Fuli previously worked at DeepSeek. Call volume surged, and during its anonymous period, total calls exceeded 1 trillion tokens, topping the OpenRouter weekly leaderboard.

On March 19 early morning, Xiaomi revealed: Hunter Alpha is actually MiMo V2-Pro. According to the same Reuters report, after the reveal, Xiaomi’s Hong Kong stock once rose by 5.8%.

This is the first time a domestic large model has proven itself on a global platform through blind testing. No branding, no publicity—just 8 days for developers to vote with their feet.