MiMo-V2.5-Pro: A 1T Open Model and What Its 1M Window Delivers

A 1M-token context window is now a common line on a spec sheet. What is less common is a vendor publishing how well its model actually reasons at that length. MiMo-V2.5-Pro is a release where Xiaomi did, and the numbers are more useful than most benchmark tables because they show the drop-off rather than hiding it.

This guide covers what V2.5-Pro is, what those long-context results tell you, how it compares with its smaller sibling, and the uncomfortable fact for anyone evaluating it in October 2026: its successor is listed at the same price.

The spec sheet

MiMo-V2.5-Pro was Xiaomi's flagship from its April 2026 release. The model card describes a sparse mixture-of-experts model:

• 1.02T total parameters, 42B active per token.
• 70 layers (one dense, 69 MoE), 384 routed experts with 8 active per token.
• Hybrid attention: 10 global layers and 60 sliding-window layers with a 128-token window.
• Three multi-token prediction layers, used for speculative decoding in Xiaomi's SGLang recipe.
• Text only, with a context window of up to 1M tokens.
• MIT licence, with a separate MiMo-V2.5-Pro-Base checkpoint (256K context) for fine-tuning.

Xiaomi positions it for agentic work, complex software engineering and long-horizon tasks with many tool calls. In the model card's benchmark table it reports 78.9 on SWE-bench Verified, 57.2 on SWE-bench Pro and 68.4 on Terminal-Bench 2.0. OpenRouter's listing also highlights strong placings on ClawEval and GDPval.

What the long-context numbers actually say

The model card reports GraphWalks results — a test where the model has to follow edges through a large graph described in the prompt — at two lengths:

• 512K tokens — 0.56 on breadth-first search, 0.92 on finding parents.
• 1M tokens — 0.37 on breadth-first search, 0.62 on finding parents.

Read that carefully. The simpler task, finding a node's parents, holds up well at 512K and loses about a third of its score at 1M. The harder multi-hop task is already weak at 512K and weaker still at full length. That is not a criticism peculiar to this model; it is what long context looks like across the field, and Xiaomi deserves credit for publishing it.

The practical lesson is that "fits in the window" and "reasoned over correctly" are different properties. If your job is locating a fact in a large document, a full window can work. If it needs chained reasoning across distant parts of the input, retrieval or staged summarisation will usually beat stuffing. Long context degradation covers the pattern in more depth.

Pro or the smaller V2.5

The MiMo-V2.5 guide covers the smaller model, a 310B-total, 15B-active omnimodal variant released alongside Pro. On Xiaomi's coding numbers the two are close: 56.1 against 57.2 on SWE-bench Pro, 65.8 against 68.4 on Terminal-Bench 2.0.

OpenRouter listed V2.5-Pro at $0.435 input and $0.87 output per million tokens at the time of writing, roughly three times the V2.5 price. That makes the decision rule fairly simple:

• Pick V2.5 when you need image, video or audio input, or when volume is high and tasks are bounded.
• Pick V2.5-Pro for long text-only sessions where a couple of points of reliability compound over many turns, and the extra active parameters earn their cost.

The successor problem

MiMo-V2.6-Pro arrived in September 2026 in the same 1.02T/42B size class, now with image, video and audio input, and OpenRouter listed it at the same price as V2.5-Pro. Xiaomi's V2.6 model card compares the two directly, and the gaps are large: 89.9 against 65.2 on Terminal-Bench 2.1, 71.9 against 19.0 on DeepSWE v1.1, and 1673 against 1107 on GDPval-AA 2.1.

Those are vendor numbers on vendor harnesses, so discount them. But a gap of that size is not going to vanish under a different scaffold. For new projects, the MiMo-V2.6-Pro guide is the more relevant read.

Running it yourself

The MIT licence makes self-hosting legally simple. Practically, a 1T model is a serious deployment. In FP8 the weights alone are on the order of a terabyte, and Xiaomi's SGLang example runs 16-way tensor and expert parallelism with EAGLE speculative decoding. Community GGUF quantisations exist for llama.cpp-based tools, which help for experimentation but not for production throughput.

Before committing hardware, run the arithmetic in the self-hosting break-even point. At an API price under a dollar per million output tokens, owning the hardware only wins at very high sustained utilisation or when data cannot leave your network.

When to keep it

V2.5-Pro is still a capable model, and there are legitimate reasons to keep running it:

1. It is already validated in production and the switch would need a fresh evaluation cycle.
2. You have fine-tuned the base checkpoint and the tuned model beats a stock V2.6-Pro on your task.
3. You self-host and your serving stack has not been updated for V2.6.

If none of those apply, move. Run your own task set on both models, compare cost per completed task, and keep V2.5-Pro as a fallback until the newer model has earned your trust.