Xiaomi's open TabLDM takes aim at the last thing LLMs never fixed: tables
Two launches this cycle push AI out of the chat window and into the two places the field has been slowest to land: structured data tables and real-time video. Xiaomi open-sourced a tabular foundation model that challenges gradient-boosting stalwarts without fine-tuning, and fal's new H3 Max renders a five-second clip in about three seconds.
Xiaomi released Xiaomi-TabLDM, an open-source foundation model built specifically for tabular data, alongside a technical report, code, and weights. The model handles classification and regression without task-specific retraining — the pitch is that teams in finance, manufacturing, healthcare, and supply chain can drop a fresh table on it and get predictions instead of rebuilding a gradient-boosting pipeline for every dataset. That is the real promise of a "foundation model" for tabular data, and it's the niche that has resisted the transformer wave longest.
The interesting engineering is how it was trained. The model saw only synthetic data generated by structural causal models, and it combines a sparse mixture-of-experts mechanism with test-time scaling, which spends more compute during inference while leaving trained parameters fixed. The benchmark results are what make the release worth a read. It took first place in regression on the OpenML-CTR23 leaderboard and topped TALENT's binary classification task, while finishing second in regression on TALENT, TabArena, and BCCO. The deployment-relevant numbers are the striking ones: on the TabArena regression benchmark Xiaomi says the model cut training time by 82% and inference time by 68% relative to TabFM, the previous top model, at only a small gap in Elo rating.
For a "foundation model" pitch the real sell is the industrial validation. In material property prediction, Xiaomi reports improved accuracy of 130% with no fine-tuning while cutting invalid sample tests by roughly 90%. On part-weight prediction it reduced error-sample rates by 31%, and in production component prediction mean error dropped 54% — then a further 62% after feeding it about 30 new samples to adapt to changed production conditions. That last figure is the punchline: the entire point of a tabular foundation model is that a couple dozen examples should pivot it to a new line instead of a month of feature engineering. It installs through pip and works with scikit-learn, so the barrier to trying it is low.
fal.ai launched H3 Max, a video-generation model built on the open-weights MiniMax H3 and retrained with an emphasis on prompt adherence and visual quality. The headline stat is speed: it renders a five-second clip in about three seconds, which fal says is roughly 35x the throughput of MiniMax's own H3 endpoint and about 15x faster than any model of comparable quality. The claim is that it holds both ends of the quality-speed curve — in fal's head-to-head human preference runs against a dozen leading video models, including Gemini Omni Flash, Kling 3, and Veo 3.1, H3 Max ranked first across overall quality, prompt understanding, and aesthetics. fal points to independent benchmarks from Artificial Analysis and Design Arena that also rank it number one.
What to watch: whether a post-trained derivative like H3 Max becomes the default way to ship video — the base model and the inference stack engineered together — and whether tabular foundation models start eating gradient boosting's market.
Which of these two lands in production faster — open tabular models or faster video generation? Tell us in the comments.
Sources: Xiaomi TabLDM (GitHub) · arXiv 2609.03880 · XimiTime · Introducing H3 Max by fal · fal H3 Max
