Xiaomi MiMo-V2.6: The Open-Weight Price Floor Just Collapsed Again

Xiaomi's MiMo-V2.6 series lands with the same price tags as V2.5, but the intelligence jump is anything but static. The defining story is not a price cut, the new models are priced identically to V2.5, but a leap in capability at a frozen price. That pushes the intelligence-per-dollar Pareto frontier out again, and it forces every other open-weight lab to respond. If you were waiting for the open-weight price war to cool off, this is the wrong week.

MiMo-V2.6 comes in three variants: Pro, Flash, and UltraSpeed. Pro is the flagship reasoning model, Flash targets low-latency applications, and UltraSpeed trades some quality for extreme speed. The headline numbers from the technical report are a 1.02 trillion parameter total with 309 billion active parameters for the Pro variant. One secondary source, deai.org, reports 524B/159B, but the vendor documentation and the primary technical report both support 1.02T/309B. Treat the larger figure as authoritative.

Pricing and the "floor" mechanic

The pricing structure for MiMo-V2.6 is aggressive, but the real story is what it does to the market. V2.5 Pro was already undercutting most proprietary rivals. V2.6 keeps the same prices while improving benchmarks across the board. That combination is a direct challenge to anyone selling intelligence at a premium.

Pro and Flash support batch pricing at 50% off. UltraSpeed does not. A May 2026 survey listed V2.5-Pro at $1.00/$3.00 versus today's $0.435/$0.87, implying a mid-year price cut, but the cut date could not be confirmed from a primary source. That is a single-source claim, so treat it as plausible but unverified.

Benchmarks: strong, but not a clean sweep

The benchmark results are the core of the argument. Xiaomi is not just matching the competition; it is beating most of them on cost-efficiency. On the Artificial Analysis Intelligence Index, MiMo Pro scores 48 on the Index, ahead of Grok 4.7 at 46, which was released the same day at $2/$6 as a proprietary model. That is a direct hit on xAI's value proposition.

• Cost efficiency: AA lists MiMo Pro at $0.13 per Index task and $207 to run the full Index, vs Qwen3.8 Max's $5.41 / $4,935.
• Pro vs. Qwen3.8 Max: MiMo Pro scores 48 on the Index, Qwen3.8 Max scores 52, but the cost gap is enormous. Pro is 41x cheaper per task.
• Flash vs. GPT-6 Astra: Flash scores 44 on the Index, trailing GPT-6 Astra's 50, but Flash costs $0.28 output vs. GPT-6 Astra's $50 short-context output. That is a 178x cost difference for a 6-point gap.
• UltraSpeed vs. Gemini 3.6 Flash: UltraSpeed scores 41, Gemini 3.6 Flash scores 47, but UltraSpeed is 5x cheaper on output.

These are not marginal differences. They are orders of magnitude. The benchmark gap to the top proprietary models is shrinking, while the price gap remains wide. That is the formula that has defined the open-weight push, and Xiaomi just executed it better than anyone else this month.

Open weights and the licensing wrinkle

MiMo-V2.6 is open-weight, but not fully open. The weights are available, but the license includes restrictions. Only Qwen3.8-27B is Apache 2.0 among the current open-weight leaders. Xiaomi uses a custom license that permits commercial use but restricts some downstream applications. That is a meaningful caveat for teams building products on top of these models.

The Flash model is the one to watch for most developers. It is priced at $0.14/$0.28, which undercuts most local models and makes it viable for high-volume inference. UltraSpeed, at $4.35/$8.70, is a niche play for latency-sensitive applications where cost is secondary. Pro sits in between, delivering near-frontier reasoning at a fraction of the cost of OpenAI's GPT-6 Astra or Anthropic's Opus 5.5.

What this means for the market

The implications for the broader market are straightforward. Xiaomi has reset the price-performance bar for open-weight models, and the pressure is now on every other lab to match it. Qwen just released Qwen3.8 Max at a much higher price point, and the benchmark gap does not justify the premium. Moonshot's Kimi K3 is the closest competitor on price, but it does not match MiMo's benchmark scores. We tracked the previous floor move when Kimi K3 broke the open-weights-cheap rule.

There are a few caveats worth flagging. The V2.5-Pro mid-year price cut is plausible but unconfirmed by a primary source. Single-source items not yet independently verified include Anthropic's Sept 10 threat report naming Xiaomi as the seventh Chinese lab, via local-ai-zone only, and Moonshot's "K2 Horizon" Apache 2.0 release in early September, via deai.org only. These should not carry load-bearing claims until corroborated. Training costs and all Xiaomi RL metrics are vendor-reported, not independently reproduced.

The takeaway is concrete: Xiaomi MiMo-V2.6 delivers near-frontier intelligence at a fraction of the cost of proprietary alternatives. The benchmark gaps to the top models are shrinking, but the price gaps remain wide. For teams building on open weights, this is the new baseline. For labs selling intelligence at a premium, the floor just got lower.

Key sources

• Xiaomi MiMo V2.6 official news
• Official pricing and comparisons
• Grok 4.7 benchmarking
• x.ai Grok 4.7 announcement