deepseek v4 is back in the spotlight after Xiaomi unveiled a new round of MiMo models on April 23 and April 24 ET, sharpening the rivalry around open models and agent tools. its Xiaomi MiMo-V2. 5-Pro reached the top tier of a major benchmark, while the team also introduced new speech models aimed at both input and output. The moves land as attention builds around deepseek v4 and the next wave of large-model competition.
MiMo V2. 5 Takes the Spotlight
On April 23 ET, Xiaomi said its self-developed Xiaomi MiMo-V2. 5-Pro ranked at the top of the open-model field on a recent Artificial Analysis list, tying for first in overall intelligence score and taking first place in the open-model Agent category. the model also placed in the top five across all large models, including closed systems, putting it alongside well-known flagship models in performance comparisons.
The timing matters. Xiaomi’s latest release adds pressure to a market already watching for deepseek v4, with the company positioning MiMo-V2. 5-Pro as its strongest model so far. Xiaomi said the model is designed for long and difficult agent tasks, with improved instruction-following and strong consistency over extended runs.
Speech Models Expand the Stack
On April 24 ET, Xiaomi’s technical team said it had added MiMo-V2. 5-TTS and MiMo-V2. 5-ASR to complete the speech input and output chain. The TTS side includes three models that can be directed through natural-language instructions to shape voice performance, including speed and emotion control. Xiaomi also said one version can generate a new voice from scratch from a single sentence, while another can clone a target voice from a few seconds of reference audio.
The ASR model was also opened with both weights and code, and Xiaomi said it is built for messy real-world settings, including Chinese dialects, mixed Chinese-English speech, strong noise, and multi-speaker dialogue. It also outputs punctuation natively so the transcript can be used immediately.
Immediate Reaction From the Company
Xiaomi said its MiMo-V2. 5-Pro is the strongest model in its MiMo line to date, and that it can match top global agent models in general intelligence, complex software engineering, and long-horizon tasks. The company also said internal testing showed the model can complete a single long task involving nearly 1, 000 tool calls when paired with the right runtime framework.
In another set of figures released by the company, MiMo-V2. 5-Pro scored 73. 7 on Xiaomi’s MiMo Coding Bench, compared with 77. 1 for Claude Opus 4. 6 and 71. 5 for MiMo-V2-Pro. Xiaomi also said the newer model improves token efficiency versus other models in benchmarked agent settings.
Why the Timing Matters
The latest rollout arrives in a tense moment for the open-model race. Xiaomi’s team is pushing beyond core language work and into agent systems, speech, and tool use, while the broader field waits for the next move in the deepseek v4 cycle. That combination has made the company’s releases hard to ignore.
For now, Xiaomi’s message is clear: the MiMo V2. 5 family is no longer a side project. With deepseek v4 drawing attention and Xiaomi expanding from text into speech and agent workflows, the next round of model competition may come fast.
