AI
Xiaomi Trained a Trillion Parameter Model Live on the Internet, and the Whole World Watched
Between September 15 and 20, Xiaomi opened its reinforcement learning stage to a live audience, setting a new precedent for training transparency. The company trained the reinforcement learning stage of its newest AI models in front of a live audience, with a public dashboard at mimo.xiaomi.com/rl ticking a cost counter upward in real time. Then it released the results as open weights under the MIT license, for anyone to download, customize, and run.
The MiMo-V2.6 family landed on September 21 and 22 with three checkpoints. Pro carries 1.02 trillion total parameters with 42 billion active per token, Flash carries 309 billion total with 15 billion active, and a 9 billion parameter distill fits on a single GPU. Every checkpoint ships with a 1 million token context window and handles text, images, video, and audio in one model. The live counters stopped at $2,620,670 for the Pro run and $854,044 for Flash, $3,474,714 combined across roughly 750,000 trajectories and 30 training steps inside six days.
The independent check came first. Artificial Analysis scored MiMo-V2.6-Pro at 46.32 on its Intelligence Index, the top open weight score on the board, roughly a point behind GPT-5.6 Sol by press coverage. Xiaomi's own model card claims near parity with Claude Opus 5 and GPT-5.6 Sol on agent benchmarks like DeepSWE and OSWorld, and the company is candid that those are vendor reported numbers, which makes the independent score the one that counts.
The livestream is the fresh angle here. Training a model has always been a closed room, a press release, and a benchmark table. Xiaomi turned it into public infrastructure, with the cost, token throughput, and benchmark curves visible to anyone with a browser. That is trust by construction, and it landed in the same week CNN reported that Chinese models grew from under 15 percent to over 54 percent of global usage in a single year, led by DeepSeek and Moonshot.
The bigger frame came from CNN's September 26 explainer on the US China AI split. President Trump told reporters he wants America to keep its AI lead because whoever wins AI wins, while Beijing counters with an open ecosystem built on accessibility and lower cost. Xi Jinping's new World AI Cooperation Organization has recruited 29 countries around open models. The quote that carries the piece comes from Chucheng Feng of Hutong Research, do I want a model that costs ten times as much but is six months ahead, or one that costs one tenth as much but is only six months behind.
For the Global South, the answer is increasingly the second one. Developers who need an inexpensive model they can download and fine tune will often take the open option over a slightly more capable system behind an expensive subscription and a foreign cloud. The Apple analogy from researcher Eric Olander sums up the possible endgame. Fewer people use Apple products, but Apple eats up 90 percent of the profits. The US could hold the profits in AI while China accumulates the users, and a trillion parameter model anyone can download is exactly how user accumulation starts.
Quick answers
What is this story about?
Between September 15 and 20, Xiaomi opened its reinforcement learning stage to a live audience, setting a new precedent for training transparency. The company trained the reinforcement learning stage of its newest AI models in front of a live audience, with a public dashboard at mimo.xiaomi.com/rl ticking a cost counter upward in real time. Then it released the results as open weights under the MIT license, for anyone to download, customize, and run.
Why does this story matter?
For the Global South, the answer is increasingly the second one. Developers who need an inexpensive model they can download and fine tune will often take the open option over a slightly more capable system behind an expensive subscription and a foreign cloud. The Apple analogy from researcher Eric Olander sums up the possible endgame. Fewer people use Apple products, but Apple eats up 90 percent of the profits. The US could hold the profits in AI while China accumulates the users, and a trillion parameter model anyone can download is exactly how user accumulation starts.
Sources
- Temperature2: Xiaomi MiMo-V2.6 open source and live RL run
- Metaverse Post: MiMo-V2.6 open source release
- eesel AI: MiMo V2.6 review and benchmarks
- CNN: The US wants the world to pick a side on AI
New to crypto? Read the crypto glossary, browse frequent questions, read our story, or explore the story archive.