# A Phone Company Just Open-Sourced the Best AI Model Money Can’t Even Buy Elsewhere

By Rafiqul Islam Rabbi · AI · Published Thu, 24 Sep 2026 18:59:57 GMT
Source: The Current Tribune — https://currenttribune.com/article/xiaomi-mimo-v2-6-open-source-ai-model-beats-deepseek

Xiaomi makes phones, rice cookers, and electric scooters. This week it also made the top-ranked open-weight AI model on one of the industry’s most closely watched scoreboards, and it did so for a training bill smaller than what some startups spend on office snacks.

## What Xiaomi Actually Shipped

On September 21, Xiaomi’s MiMo team released two new models under the fully open MIT license: MiMo-V2.6-Pro and a smaller sibling called MiMo-V2.6-Flash, alongside a faster-inference variant called Pro-UltraSpeed. Both are mixture-of-experts models, meaning they carry a huge number of parameters in storage but only activate a fraction of them for any given task — the trick that makes trillion-parameter models practical to run at all.

Pro is the headline number: 1.02 trillion total parameters, with 42 billion active during inference. Flash is the leaner option at 310 billion total parameters and 15 billion active. Both share a 1-million-token context window, can take text, image, audio, and video as input, and can generate up to 128,000 tokens of output in a single response. Xiaomi didn’t just drop the weights and walk away, either — it published full technical documentation, more than 7,000 reinforcement-learning training environments, and a distilled 9-billion-parameter version small enough to run on considerably more modest hardware.

### Where It Lands on the Leaderboard

On Artificial Analysis’s Intelligence Index, a benchmark aggregator that’s become something like the Billboard chart of AI capability, MiMo-V2.6-Pro scored 46 — tying xAI’s Grok 4.7 and edging out Grok 4.6 (44) and Google’s Gemini 3.8 Flash (41). DeepSeek, the Chinese lab that spent most of this year as the open-weight benchmark to beat, trails clearly here: its V4.1 Flash scores 39 and V4.1 Pro scores 36.

The bigger jump shows up in agentic coding benchmarks, where the gains from the previous MiMo generation are almost startling. On DeepSWE v1.1, a benchmark that tests whether a model can actually complete real software engineering tasks rather than just describe them, Pro jumped from 19.0 to 71.9 — a nearly fourfold improvement in a single generation. Flash isn’t far behind at 67.9, which is remarkable for a model less than a third of Pro’s size.

## It’s Not Beating Everyone

Credit where it’s due, but this isn’t a story about Xiaomi dethroning the entire industry. Anthropic’s Claude Opus 5 still leads MiMo-V2.6 on several evaluations, including that same DeepSWE benchmark and the newer Terminal Bench 4.0, which tests a model’s ability to operate inside a real command-line environment over long, multi-step tasks. Closed frontier labs still hold the ceiling on the hardest agentic work. What Xiaomi has done is close the gap between best open model and best model, period, to a degree that would have seemed implausible a year ago.

### The Price Tag Is the Real Headline

Here’s the number that should worry every AI lab charging premium API rates: MiMo-V2.6-Pro costs $0.435 per million input tokens and $0.87 per million output tokens, with cached input dropping to a startling $0.0036 per million. Flash runs even cheaper, at $0.14 and $0.28 per million respectively. That’s roughly a third to a tenth of what several closed frontier models charge for comparable performance on agent tasks.

The training cost behind it is arguably the wilder number. Xiaomi says Pro and Flash went through 30 reinforcement-learning steps covering roughly 750,000 trajectories, completed in six days, at a total cost of $2.62 million for Pro and $850,000 for Flash. Labs have spent more than that on a single week of compute for a model a fraction as capable. The method behind it — an asynchronous variant of Group Relative Policy Optimization, with custom reward shaping specifically designed to stop the model from gaming its own training signal — is the kind of engineering detail that usually stays buried in a technical appendix. Here, it’s the whole story.

## Why a Phone Maker Is Doing This at All

Xiaomi’s AI ambitions aren’t a side project. The company has been steadily building MiMo as the intelligence layer behind its own devices and, increasingly, as a credibility play in a market where having an in-house AI lab has become table stakes for any consumer electronics giant. Open-sourcing a trillion-parameter model that beats DeepSeek on public benchmarks is as much a recruiting and reputation move as it is a product release — it tells enterprise customers and researchers that Xiaomi’s AI division isn’t an afterthought bolted onto a smartphone roadmap.

It also puts pressure on the rest of the Chinese AI field, where DeepSeek, Alibaba’s Qwen, and Moonshot have all been racing to define what open and competitive with the West actually looks like. MiMo-V2.6 resets that internal competition, and it does it with numbers that are easy to verify since the weights are sitting on Hugging Face for anyone to download and test themselves.

## What This Means

MiMo-V2.6 isn’t the best AI model in the world. It’s the best one you can download, modify, and run on your own infrastructure without asking anyone’s permission — and it got there at a training cost that makes the entire billions-of-dollars-just-to-compete narrative around frontier AI look less like a law of physics and more like a choice some labs are making. For developers and companies deciding where to build, that distinction matters more than a two-point gap on a leaderboard. Watch what Alibaba and DeepSeek do next; this kind of price-performance pressure rarely goes unanswered for long.