Home / Blog / V2.5 Series Deep Dive

Published: April 29, 2026 · Updated: July 2026

MiMo-V2.5 Series: Pro · Omni · TTS — A Complete Breakdown

Overview

In April 2026, Xiaomi open-sourced three new models under the MIT license, representing the most significant expansion of the MiMo family since V2-Flash. The V2.5 series targets three distinct use cases:

• MiMo-V2.5-Pro: Agent-optimized flagship with 1M context window. Designed for complex multi-step reasoning tasks, tool use, and long-document analysis.
• MiMo-V2.5-Omni: Full-modal model capable of processing and generating across text, image, video, and audio modalities.
• MiMo-V2.5-TTS: Speech synthesis with natural prosody, supporting bilingual (Mandarin + English) output and multiple Chinese dialects.

MiMo-V2.5-Pro: The Agent Flagship

What's New

• 1M context window — 20× larger than V2-Flash (56k). Achieved via FlashAttention-3 + ring attention for distributed inference.
• 1T+ total parameters — exact architecture hasn't been disclosed, but inference requires approximately 4× A100-80GB nodes in FP16.
• Agent-optimized training — fine-tuned on tool-use trajectories, multi-turn dialogues, and long-form reasoning chains.
• MIT licensed — same as all other MiMo models.

API Pricing

Permanently reduced May 2026:

• Input: $1 per million tokens
• Output: $3 per million tokens

Benchmarks

MiMo-V2.5-Pro focuses on agentic and long-context tasks. While SWE-Bench and AIME numbers haven't been published for this model (it's designed for different workloads), its strength is sustained multi-turn reasoning across very long contexts. In internal evaluations, it maintains coherent reasoning over 500K+ token documents, significantly outperforming V2-Flash on long-context retrieval and synthesis tasks.

MiMo-V2.5-Omni: Full-Modal Understanding

V2.5-Omni processes text, images, video, and audio in a unified architecture. Unlike pipelines that chain separate vision and language models, Omni uses a shared transformer backbone with modality-specific encoders:

• Vision: ViT-based encoder with dynamic resolution (supports up to 4K images)
• Video: Temporal attention module processing up to 60 seconds at 2fps
• Audio: Whisper-style encoder for speech, plus a general audio encoder for music/sound
• Text: Standard token embedding with the MiMo reasoning backbone

All modalities feed into a shared reasoning layer, allowing cross-modal understanding — for example, analyzing a video of a cooking process and answering questions about ingredient timing. MIT licensed.

MiMo-V2.5-TTS: Speech Synthesis

V2.5-TTS is a standalone speech synthesis model with:

• Bilingual output (Mandarin Chinese and English) with code-switching support
• Multiple Chinese dialects (Cantonese, Shanghainese, Sichuanese — in beta)
• Voice cloning from 10-second reference audio
• Emotional prosody control via text tags
• Real-time inference on GPU (sub-200ms latency for 5-second utterance)

The model is trained on a proprietary corpus of 50K+ hours of labeled speech data. Xiaomi has indicated commercial licensing terms for TTS use cases will be published separately from the model weights.

How V2.5 Fits the MiMo Lineup

Model	Parameters	Context	Best For	Pricing
MiMo-7B	7B	32k	Edge, on-device	Free (open weights)
V2-Flash	309B/15B active	56k	High-speed reasoning, coding	$0.50/$1.50 per M
V2.5-Pro	1T+	1M	Agentic, long-context	$1/$3 per M
V2.5-Omni	TBD	TBD	Multimodal understanding	TBD (API) / MIT weights
V2.5-TTS	TBD	N/A	Speech synthesis	Commercial license
MiMo Code	Agent	Infinite	AI coding agent	Free app + API usage

Availability

• V2.5-Pro: Available via API at platform.xiaomimimo.com and as open weights on HuggingFace
• V2.5-Omni: Open weights on HuggingFace; API tier coming
• V2.5-TTS: Open weights on HuggingFace; commercial license required for production voice cloning

All weights are at huggingface.co/XiaomiMiMo.