Realtime AI News
Xiaomi closes six-day livestreamed training run with over 20 million yuan spent and the open-source top spot
Xiaomi turned a six-day model training run into a public livestream that cost more than 20 million yuan, according to QbitAI, and finished with the top spot among open-source models. Hugging Face's CEO publicly called it 'great,' making the run an unusually transparent moment in the open-source race.
A six-day training experiment by Xiaomi, billed as a 'training livestream,' wrapped up and was reported on September 22 by the Chinese outlet QbitAI. The whole run was shown to viewers from the start of training to the final settlement, and the bill came to more than 20 million yuan.
The result was striking as well: once training finished, Xiaomi's model held the top spot among open-source models, and Hugging Face's CEO publicly praised the run as 'great,' the report says.
Broadcasting an entire training run is itself a statement. Open-source communities usually see only the final weights and a technical report; how much compute gets burned and how a run is driven forward day by day normally stays inside the team. Here the cost and the process were exposed together.
For the open-source ecosystem, the more interesting change is positional. The gap between open-source models and the closed frontier has been debated for two years. The top of the open-source table is no longer just a technical ranking; it shapes which weights downstream companies treat as their default base.
Xiaomi's entry adds a player from consumer electronics and smart hardware to the race. Companies like this have devices, user scenarios and a direct channel into hardware and system software, so choosing to open-source means releasing part of that capability to the wider community.
The other side of the story is cost. A bill above 20 million yuan for six days is a blunt reminder that competitive frontier training remains expensive, and any open-source effort that wants to iterate at this intensity has to answer where the compute comes from and whether the spending can be sustained.
The concrete things to watch next: how open the weights and licence turn out to be, whether third parties can reproduce the gains from this run, and whether the approach grows from a livestream experiment into a durable model service and product capability.
Why it matters
By livestreaming a training run worth tens of millions of yuan, Xiaomi pushes the open-source race from publishing weights toward publishing cost and process, raising expectations for transparency and giving outsiders a clearer yardstick for the distance between open-source leaders and the closed frontier.
Nearby Updates
All09/22, 19:58
WebArena author Shuyan Zhou joins Meta Superintelligence Labs to build an AI browser
Shuyan Zhou, the author of the WebArena benchmark for web agents, has joined Meta's Superintelligence Labs, where her core focus will be building an AI browser. She has said the research interest traces back to a simple goal: a device that can book flight tickets for her mother through conversation.
09/22, 20:00
NVIDIA Releases Isaac ROS 5.0 to Push Agentic, Open Source Robotics
NVIDIA has released Isaac ROS 5.0, a collection of GPU-accelerated packages built on ROS for developers who build robotics applications that perceive, reason and act in dynamic environments. The company says such work needs new physical AI models and tools, and the release keeps the stack open source.
09/22, 20:23
Nscale's IPO will test Wall Street's appetite for concentrated AI bets once again
Nscale, a British AI data center developer, is heading to public markets with most of its revenue coming from Microsoft and Anthropic. The listing will test whether public investors are still willing to pay up for AI infrastructure bets concentrated on a very small number of customers.
09/22, 18:06
Jiyuan Ludong CTO Han Kai at Yunqi Conference: From Multi-Model Routing to an Agent Feedback Loop
On September 22, Jiyuan Ludong co-founder and CTO Han Kai spoke at the enterprise agent summit of the 2026 Yunqi Conference in Hangzhou, in a talk titled From Harness to RSI Flywheel. The session focused on moving from multi-model routing to a feedback loop that lets enterprise agents keep improving.