Xiaomi Just Dropped MiMo-V2.6-Pro — an Open-Weights Trillion-Parameter Model That Tops the Intelligence Charts

Introduction

Every few months the AI industry gets a reminder that the frontier is no longer the exclusive property of a handful of US labs. Today, that reminder came from an unexpected place: a smartphone company. Xiaomi has officially released MiMo-V2.6-Pro, the latest and largest entry in its MiMo model family, and the numbers attached to it are turning heads across the industry.

MiMo-V2.6-Pro is a mixture-of-experts (MoE) model with roughly 1 trillion total parameters, of which only about 42 billion are active per token. It ships under a permissive MIT license, meaning anyone can download the weights from Hugging Face, run them commercially, fine-tune them, or build products on top of them — no per-seat negotiations, no API-only gatekeeping.

And unlike many "open" releases that quietly trail the closed frontier, this one competes at the top: MiMo-V2.6-Pro scores 46 on the Artificial Analysis Intelligence Index, placing it first among the 114 models in its evaluation class, while serving tokens at a brisk ~125 tokens per second. Its recommended retail pricing is $0.435 per million input tokens and $0.87 per million output tokens — a fraction of what closed flagship models charge.

Here's everything you need to know, and why this launch matters more than the benchmark chart suggests.

The Headline Specs

Let's start with the raw numbers, because they're unusual in combination:

• Total parameters: ~1.0T (the full size of the network)
• Active parameters: ~42B per token (what actually fires during inference)
• Architecture: Sparse mixture-of-experts (MoE)
• Context window: 1,000,000 tokens — roughly 1,500 A4 pages of 12pt Arial text
• Input modalities: Text, image, speech, and video
• Output modality: Text
• License: MIT (fully open, commercial use allowed)
• Speed: ~124.5 output tokens per second on standard serving hardware
• Cache discount: 99% on repeated/cached input tokens ($0.13 per 1M cached tokens)
• Intelligence Index: 46 (Artificial Analysis), #1 of 114 in its class

That combination — trillion-parameter scale, tiny active footprint, million-token context, omni-modal input, and a fully permissive license — is genuinely novel. Most open-weights releases force you to pick two or three of those attributes. MiMo-V2.6-Pro ships all of them at once.

Why "42B Active Out of 1T" Matters

If you're new to mixture-of-experts architectures, here's the short version: instead of one monolithic neural network that processes every token with all of its weights, an MoE model contains many specialized "expert" sub-networks. A lightweight router looks at each token and decides which handful of experts should process it.

The practical consequence is that MiMo-V2.6-Pro gets something close to the knowledge capacity of a trillion-parameter model while paying the compute cost of a 42B-parameter model on every single token. That's why it can generate at 125 tokens per second — speeds that dense models of comparable total size simply cannot reach on the same hardware.

There are trade-offs, of course:

• Memory footprint: You still need to hold all ~1T parameters in (high-bandwidth) memory, even though only a slice is used per token. Hosting MiMo-V2.6-Pro yourself requires serious infrastructure — think multi-node GPU clusters, not a single gaming rig.
• Routing overhead: Expert routing adds a small per-token cost and, in some deployments, can create hotspots if the router sends disproportionate traffic to particular experts.
• Verbosity: Early evaluations note the model is somewhat more verbose than the median — it "thinks out loud" more, which matters if you're paying per output token at scale.

Still, for most buyers the equation is simple: frontier-adjacent quality at mid-tier pricing, with the option to self-host.

The Benchmark Picture

Independent evaluation data published today shows MiMo-V2.6-Pro at the top of its class:

Two things stand out. First, the quality gap versus the median open-weights model is enormous — 46 versus 18 means this is roughly twice as capable as the "typical" model in its peer group. Second, despite being priced slightly above the median on input tokens, it's cheaper than the median on output tokens — and output tokens are usually where the bill accumulates, especially for reasoning-heavy workloads where models generate long chains of thought.

The 99% cache discount is the sleeper feature here. If your application re-sends a large, stable system prompt or a long document corpus with every request (RAG systems, coding agents, customer-support bots), cached input tokens at $0.13 per million can cut your effective input bill by an order of magnitude.

What "MIT License" Changes for Builders

The open-weights landscape in 2026 is a patchwork. Some models ship "open weights, non-commercial." Some ship under community licenses with acceptable-use riders. A handful ship under true permissive licenses. MiMo-V2.6-Pro's MIT license puts it in the most builder-friendly category:

1. Commercial products — you can wrap it in a product, charge for it, and never owe royalties.
2. Fine-tuning and distillation — you can specialize it on your data and even ship derivative weights.
3. No usage telemetry — nothing phones home; what you run on your cluster stays on your cluster.
4. Downstream certainty — MIT is legally boring, which is exactly what corporate legal teams want to hear.

For startups that were renting flagship API access at $5–$25 per million output tokens, a model with 90%+ of the capability at under $1 per million output tokens — that they could also self-host — changes build-versus-buy math dramatically.

The Chinese Open-Weights Wave Continues

MiMo-V2.6-Pro doesn't arrive in a vacuum. Over the past year we've seen DeepSeek, Zhipu (GLM), Moonshot (Kimi), and Qwen each push open-weights models into territory that was recently closed-model-only. The pattern is consistent:

• Aggressive MoE efficiency — huge total capacity, small active compute.
• Long context as a default — 256K to 1M tokens rather than premium add-ons.
• Omni-modality creeping in — text plus vision, and increasingly audio and video input.
• Permissive licenses — removing the last excuse for enterprises to stay on closed APIs.

Xiaomi entering this race at trillion-parameter scale says something important: the expertise and capital required to train a frontier-class model has diffused far beyond the original AI labs. A company whose core business is phones and IoT devices now fields a model that tops independent intelligence rankings. Expect more consumer-hardware giants to follow.

What This Means for Pricing Across the Market

Competition from open-weights models has been the single biggest deflationary force in AI pricing. When a model scoring 46 on the Intelligence Index costs $0.87 per million output tokens, it puts immediate pressure on closed models charging $5–$25 for comparable scores. We saw this movie with DeepSeek V4 earlier this year, and each round of open releases resets buyer expectations lower.

For developers, the practical advice is to re-benchmark your workloads quarterly. The model that was the obvious choice six months ago may now be 10–20x more expensive than an equally capable alternative. If you're routing between models, make sure your router can actually pick up new releases like this one on day one — which is exactly why we keep Qubax's model catalog updated in near-real-time as launches land.

How to Try MiMo Today

The weights are on Hugging Face under MIT. If you don't want to stand up a multi-GPU cluster to serve a trillion-parameter MoE yourself, you can call MiMo-family models through an OpenAI-compatible API in a few lines of Python:

from openai import OpenAI

client = OpenAI(
    base_url="https://api.qubax.ai/v1",
    api_key="YOUR_QUBAX_KEY",
)

resp = client.chat.completions.create(
    model="mimo-v2.6-pro",
    messages=[
        {"role": "system", "content": "You are a precise research assistant."},
        {"role": "user", "content": "Summarize the trade-offs of MoE vs dense models."},
    ],
)
print(resp.choices[0].message.content)

Browse pricing for MiMo and every other model on the Qubax models page, and check the API docs for streaming, tool-calling, and structured output support.

The Bottom Line

MiMo-V2.6-Pro is the strongest signal yet that open-weights models aren't chasing the frontier — in some measurable dimensions, they are the frontier. A trillion parameters, 42B active, 1M context, omni-modal input, MIT license, and sub-dollar output pricing is a package that would have sounded absurd eighteen months ago.

The interesting question for the next quarter isn't whether open models catch up to closed ones. It's how fast closed-model pricing has to fall to stay in the conversation.

Try MiMo-V2.6-Pro and every other frontier model on Qubax →

FAQ

Is MiMo-V2.6-Pro really free to use commercially?

The model weights are MIT-licensed, so commercial use, modification, and redistribution are all permitted. If you self-host, your costs are infrastructure only. If you use a hosted API, you pay per token — typically a small fraction of closed-model pricing.

How much hardware does it take to run MiMo-V2.6-Pro locally?

Although only ~42B parameters activate per token, all ~1T parameters must be resident in memory, so self-hosting requires a multi-node high-bandwidth GPU cluster. For most teams, hosted inference is far more economical than owning that hardware.

How does MiMo-V2.6-Pro compare to GPT-6 Astra or Claude Opus 5?

On the Artificial Analysis Intelligence Index it scores 46, placing it at the top of its class, at roughly 1/10th to 1/30th the output token price of those closed flagships. Closed models may still lead on specific agentic or ecosystem-integration features, so benchmark against your own workload before switching.

Does it support tool calling and structured output?

Yes. MiMo-V2.6-Pro supports reasoning modes and is compatible with OpenAI-style tool-calling APIs, so existing agent frameworks can use it with minimal changes. See the Qubax docs for details.

What is the context window and cache pricing?

1 million input tokens, with a 99% discount on cached input tokens ($0.13 per 1M vs $0.435 uncached). Long, stable prompts — system prompts, codebases, document sets — become dramatically cheaper when caching is enabled.