MiMo-V2.5: Xiaomi's Omnimodal MoE at 15B Active Parameters

The usual assumption with a two-tier model family is that the smaller one is the compromise: cheaper, weaker, fine for drafts. MiMo-V2.5 is a case where that assumption is only half right. It is cheaper and smaller than its Pro sibling, but on Xiaomi's own coding numbers the gap is one to three points, and it does something the Pro model cannot — it takes images, video and audio as input.

Xiaomi released both in April 2026 under the MIT licence. Since then the V2.6 generation has arrived, so the question for anyone evaluating V2.5 today is narrower: what does it do, and is there still a reason to pick it over its successor?

What is in the box

MiMo-V2.5 is a sparse mixture-of-experts model with 310B total parameters and 15B active per token, built on the MiMo-V2-Flash backbone. The model card gives the details:

• 48 layers, one dense and 47 MoE, with 256 routed experts and 8 active per token.
• Hybrid attention: 9 global layers and 39 sliding-window layers with a 128-token window.
• A 729M-parameter vision encoder and a 261M-parameter audio encoder, so image, video and audio are native inputs. Output is text.
• Three multi-token prediction layers, which serving engines can use for speculative decoding.
• Pre-training on roughly 48T tokens in FP8, and a context window extended in stages to 1M tokens.

The 15B active figure is the one that sets your running cost. It is a small forward pass for a model in this class, which is why the per-token price is low. Mixture of experts explained covers why total and active parameters measure different things.

How close it gets to Pro

OpenRouter's listing describes V2.5 as delivering Pro-level agentic performance at roughly half the inference cost. Xiaomi's model cards let you check that claim on two coding benchmarks:

• SWE-bench Pro — 56.1 for V2.5, 57.2 for V2.5-Pro.
• Terminal-Bench 2.0 — 65.8 for V2.5, 68.4 for V2.5-Pro.

Those are vendor-run figures and should be treated as such, but the shape is clear: on agentic coding the smaller model sits within a few points of the flagship. The V2.5 card also reports Claw-Eval results: 62.3 on the text split and 23.8 on the multimodal split. No model in Xiaomi's multimodal chart passes 26 on that second split, a reminder that agentic tasks over images remain hard across the field.

On price the gap is wider than half. OpenRouter listed V2.5 at $0.14 input and $0.28 output per million tokens at the time of writing, against $0.435 and $0.87 for V2.5-Pro — closer to a third. The MiMo-V2.5-Pro guide covers what the larger model buys you, which is mostly depth on long text-only tasks.

The window is 1M, the attention is mostly local

In V2.5, 39 of the 48 attention layers look only at a 128-token sliding window; the other 9 are global. That is what makes a 1M-token context affordable to serve: the KV cache for sliding-window layers stays small no matter how long the prompt gets, and only the global layers carry the full history.

The trade-off is the usual one. A model can accept a million tokens without using all of them equally well, and Xiaomi's own SGLang example for V2.5 sets a 256K context length rather than the full 1M. Test retrieval and reasoning at the lengths you actually plan to use. Long context degradation explains how quality tends to fall before the advertised limit.

MIT weights and what self-hosting looks like

The weights are on Hugging Face under MIT, alongside a MiMo-V2.5-Base checkpoint with a 256K context for anyone who wants to fine-tune from a pre-trained starting point. MIT means no revenue threshold and no separate commercial agreement.

Self-hosting is plausible rather than easy. In FP8 the full model is roughly a byte per parameter, so expect around 300 GB of weights before KV cache. Xiaomi's SGLang example uses eight-way tensor parallelism. The community has also published GGUF quantisations on Hugging Face for llama.cpp-based tools, which lowers the bar for experimentation considerably.

One practical warning from the model card: Xiaomi updated config.json and tokenizer_config.json after the initial release. If you downloaded early, re-pull those files, because the outdated config can degrade output quality.

V2.5 now that V2.6-Flash exists

MiMo-V2.6-Flash is the direct successor: the same size class, the same omnimodal inputs, and the same OpenRouter price at the time of writing. Xiaomi's V2.6 benchmark table compares Flash against V2.5-Pro rather than against V2.5, so there is no published head-to-head, but the scale of V2.6's gains over V2.5-Pro makes it very likely the newer model is stronger for agentic work. The MiMo-V2.6-Flash guide has the numbers.

There are still reasons to stay on V2.5:

• You have already validated it on your workload, and re-validation costs more than the expected gain.
• You self-host and depend on a quantised build or serving recipe that has not caught up with V2.6 yet.
• You need behaviour that does not move while you run a long comparison.

Otherwise, for new work, start with V2.6-Flash and keep V2.5 as a known-good fallback.

Takeaway

MiMo-V2.5 is a strong small-active-parameter model with native multimodal input, a permissive licence and coding scores close to its own flagship. Its main weakness in October 2026 is simply that a newer sibling exists at the same price. If you are choosing today, test both on ten of your own tasks, and pick V2.5 only if it wins or if stability matters more to you than the upgrade.