Mi and Mini Max both zoomed in

On 18 and 19 March, two Chinese companies released their respective Large Agent Direction Models. The country's AI Creative Corporation Mini Max launched M2.7, and Mimo launched V2-Pro. Both models entered the first global ladder on Agent benchmark, but their API output pricing was 1/21 and 1/8 respectively of Claude Opus 4.6。

They're playing cards the same week, but they have different cards. They represent two distinct technological paths, and they bet on the future of the Age。

Same test, 1/17 tuition fees

Look at the most intuitive comparison。

According to OpenRouter and official pricing pages of companies, MiniMax M2.7 is US$1.2 and MiMo-V2-Pro is US$3 at API output prices (per million tokens). For reference purposes, Claude Opus 4.6 has an output price of $25, GPT-5.2 is $14 and Claude Sonet 4.6 is $15。

The price gap is quantitative, but not capacity. On SWE-bench Verified (the most mainstream of currently measured code engineering capabilities in benchmark), MiMo-V2-Pro received 78 per cent, Sonet 4.6 was 79.6 per cent, a difference of less than two percentage points. The SWE-Pro achievement of M2.7 was 56.22 per cent, the same as GPT-5.3-Codex. On VIBE-Pro (end-to-end project delivery capability), M2.7 received 55.6 per cent, close to the level of Opus 4.6。

The focus of this picture is not on who is who is who is who is who is who is who is who is who is who is who is who is who is who is who is who is who is who is who is who is who is who is who is who is who is who is who is who is who is who is who is who is who is who is what is who is who is who is who is who is who is who is who is who is who is who is who is who is who is who is who is who is who is who is who is who is who is who is who is who is who is who is who is who is who is who is who is who is who is who is who is who is who is who is. The point is that the price-performance of the scissors is poor: the Agent model has been squeezed into the same power belt, but is in a completely different price zone。

hundreds of billions of parameters vs evolution

Prices are just symptoms. The two companies produced two completely different sets of bottom cards。

MiMo-V2-Pro is following the "Mixed Miracle" route. According to the rice official announcement, V2-Pro has more than 1 trillion total parameters, activates 42B and supports the extra-long context of 1 million tokens. Its core innovation is the Hybrid Attention Mixed Attention Mechanism, which brings the ratio of sliding window attention (SWA) to global attention (GA) to 7:1 — V2-Flash is 5:1. This set-up makes the model more stable when dealing with the Agent scene in parallel with long files and multiple tools. On PinchBench (Agent Tool Callability Assessment), MiMo-V2-Pro received 84%。

M2.7 A completely different path was taken. According to the official technical blog published by Mini Max on March 18, the parameters of M2.7 are not publicly available, but it shows a mechanism of "self-inhibitive evolution": models operate autonomously for more than 100 rounds of optimized cycles, including analysis of failed tracks, planning modifications, modification of their own code structure, running assessments, recycling, and ultimately 30% performance improvement in internal assessment sets. Of the 22 high-cost questions of MLE Bench Lite, M2.7 took 9 gold 5 silver 1 copper, with an average medal rate of 66.6 per cent。

In terms of five dimensions, the two routes are headed in a completely different direction: MiMo-V2-Pro has a clear advantage over context length and code engineering dimensions, and M2.7 has a distance between office automation and iterative capacity. According to the same technical blog of Mini Max, M2.7 obtained ELO 1495 on GDPval-AA (Office Documents Processing Assessment), first in the open source model, and maintained a 97% skills compliance rate in the MM-Claw test covering over 40 complex skills。

Five-month, four

The two companies have not only different technical routes, but also completely different pacing。

According to public release records, Mini Max published M2 in October 2025 and M2 in March 2026, four versions in five months, one large version per 49 days. The interval between M2.5 and M2.7 is about 30 days。

MiMo ' s rhythm is different: MiMo-7B was released in April 2025 (an open source reasoning model for 7B parameters), V2-Flash was released in December of the same year (309B total parameters) and V2-Pro (1T total parameters) was released in March 2026. The size of the parameters varies from generation to generation, but the intervals between versions are longer。

Mini Max has chosen a small run, with a small but very high frequency each time, and M2.7's mechanism is itself designed for "continuous evolution". Mi chose to take a strong shot, and each version represented a significant leap in the size and structure of the parameters。

Anonymous 8 days, top of the top OpenRouter

In addition to technical routes, the marketing strategy for millet also breaks industry practices。

Reuters reported that on March 11, an anonymous model named Hunter Alpha appeared on the largest API aggregate platform in the world, OpenRouter. No brand endorsement, no press conference, no technical blog. It's extremely low-priced, and it's surprisingly strong。

The community began to speculate about its origins. According to Recublic World and numerous scientific and technological media reports, the most dominant guess is DeepSeek V4, since the head of the Mimo team, Roosevelt, had previously done research in DeepSeek. The number of calls rose rapidly, and the total number of calls exceeded 1 trillion tokens during the anonymous period, peaking at the top of the OpenRouter week list。

In the early hours of March 19th, Xiaomi revealed that Hunter Alpha was MiMo-V2-Pro. According to the same report by Reuters, after the announcement, the Mi Harbour stock rose by 5.8 per cent。

This is the first time that a large model of national production has proved itself on a global platform in a purely blind manner. Not by brand, not by publicity, eight days for developers to vote。