Mitra-v2: Synthetic-Data-Only 77M Tabular Model Matches 1.6B Rivals on TabArena
chaumian · x · 2026-09-08
Mitra-v2 Technical Report
- What it is: A tabular foundation model for real-world classification and regression — credit-risk scoring, clinical prediction, equipment-failure detection, house-price estimation.
- Trained purely on synthetic data, with a far larger and more diverse pretraining task distribution than Mitra-v1; built on a small 2D Transformer backbone supporting longer contexts and larger feature spaces.
- Results: Evaluated on TabArena and TALENT benchmarks spanning 300+ real datasets. It matches industry-scale TabFM and EXAONE Tabular at SOTA level and beats TabPFN-3 by a wide margin on both classification and regression.
- Efficiency: With only 77M parameters (5% the size of the 1.6B TabFM), it delivers frontier performance at a fraction of the cost.
- Generalization: Ranks first on classification tasks with 10+ classes despite being pretrained only on tasks with at most 10 classes; one of the strongest open tabular foundation models.
More from Research
- Tao weighs in as Anthropic rumored to have solved a Millennium Prize Problem — lukaszkaiser · 2026-09-08
- Lab builds Physical AI from pure logic gates, banning floats and neural nets — cephaloform · 2026-09-08
- Six ECCV papers from one team: multimodal, agents, 3D and embodied AI — liuziwei7 · 2026-09-08
- A mathematician's deep dive into Lean: tactics make formal proofs hard to read — burny_tech · 2026-09-08
- Motus2 world model hits 84% success on dexterous tasks with self-evolving closed loop — burny_tech · 2026-09-08
- Mathematician: Astra solved two of my unpublished theorems in ~60 hours each, $20k in API — basedjensen · 2026-09-08