🍭 The Best Open Model Is Now Xiaomi 🟧

Good morning. The company that makes your rice cooker now makes one of the best AI models on the planet, and it's free.

Let’s dive in 👇

🍳 What’s Cookin’:

• Xiaomi now makes the best open model in the world. Yes, that Xiaomi.

• Alibaba unveiled China's most powerful AI chip.

• Could frontier labs install a kill switch, please? Asking for Newsom.

Xiaomi
🟧 The $2.6M Model That Tied Grok

The Bite:

Xiaomi released its MiMo-V2.6 model family.

MiMo-V2.6-Pro scored 46 on Artificial Analysis's Intelligence Index, the highest mark any open-weight model has posted.

That ties it with xAI's Grok 4.7, which launched the same day.

The weights are on Hugging Face under an MIT license, free to download, fine-tune, and self-host.

Xiaomi says the reinforcement learning run behind Pro cost about $2.62 million.

Claude Opus 5 and GPT-5.6 Sol still lead on several of the benchmarks Xiaomi published.

Snacks:

• MiMo-V2.6-Pro is a 1.02-trillion-parameter mixture-of-experts model with 42 billion parameters active per token.

• Pro costs $0.435 per million input tokens and $0.87 per million output. Claude Opus 5 costs $5 and $25.

• Flash trails Pro by roughly two points on Terminal Bench 2.1 at about a third of the price.

• Xiaomi also published 7,000+ reinforcement learning environments, the training framework, and a distilled 9-billion-parameter model.

• Pro and Flash each ran 30 reinforcement learning steps across roughly 750,000 trajectories in under six days.

• Confirmed reward-hacking runs accounted for less than 2%, after Xiaomi built a dedicated "hack agent" to hunt loopholes before training.

Why it Bites:

A consumer electronics company just took the top open-weight spot on Artificial Analysis, with about $2.6 million in reinforcement learning compute.

Compared to the frontier private models, the capability gap is a few benchmark points, but the cost gap is two orders of magnitude.

That's the leverage moving.

And it moves further because of what Xiaomi shipped next to the weights: the environments, the training framework, the graders, and its defenses against reward hacking.

Releasing a competitive model pressures pricing for a quarter.
Releasing the recipe pressures it indefinitely.

So the buying question isn't whether MiMo matches Opus 5. It doesn't, quite. The question becomes whether Flash at $0.28 per million output tokens is close enough for the 90% of agent calls that never needed a frontier model.

The frontier labs still own the ceiling. They've stopped owning the floor.

Steal This Prompt
🧶 Cozy Crochet Tales

Turn your ideas into tiny, ridiculously cute crochet characters that look handmade, tactile, and ready for their own little storybook.

Use it to:

• Turn characters into handmade crochet dolls

• Build tiny woodland and cottagecore scenes

• Give animals cozy jobs, outfits, and miniature props

Workflow:

1. Hit this link: Cozy Crochet Tales

2. Paste into your AI model

3. Replace the #’s with your character or scene

   Try on Snack Prompt

_ToolBox™
🧰 5 BRAND NEW AI LAUNCHES

🔍 RankControl

Writes and publishes on your own domain, then tracks whether ChatGPT, Perplexity, Claude, and Gemini actually cite you, plus where competitors are getting picked instead.

🌍 Linguo Translate

Lives in your macOS menu bar, translates 22 languages fully offline, and swaps to Apple Intelligence, Gemini, OpenAI, or Anthropic when the offline version isn't good enough.

🛡️ Koreshield

Screens the customer message, the retrieved doc, and the proposed tool call before your support agent acts on any of them, and logs every decision for when someone asks what happened.

🤖 Solid

Agents that provision their own machines, open their own accounts, and pay for their own services, which is either the future of work or a procurement nightmare.

🎬 Pactto

Persistent rooms where creative review actually sticks, so feedback stops dying in Slack threads and call recordings nobody rewatches.

_{Everything Else}
🧠 You Need to Know

🔌 Alibaba Unveils China's Most Powerful AI Chip
→ Chief executive Eddie Wu introduced the Zhenwu V900 at Tuesday's Apsara Conference in Hangzhou, claiming three times the performance of its predecessor, with mass production targeted for the first quarter of 2027.

📊 Most Americans Say AI Firms Aren't Doing Enough
→ A Reuters/Ipsos poll of 1,277 US adults found 73% think AI companies have not gone far enough to prevent serious harm, and 55% would welcome slowing development.

🏛️ Newsom Orders Work On An AI Kill Switch
→ The California governor signed an executive order giving a panel of experts two months to recommend stronger state AI rules, including an emergency shutoff requirement for frontier models.

💳 Banks Warn Against AI Shopping Agents
→ NatWest, Bank of America and four other lenders published a report saying agentic commerce is advancing faster than industry standards and consumer protections.

🧠 Xiaomi Tops Open Weight Models With MiMo-V2.6
→ Xiaomi released the MiMo-V2.6 family under an MIT license Monday, with Pro scoring 46 on Artificial Analysis's Intelligence Index, the highest of any open-weight model.