{"passport":{"slug":"tabpfn-3","display_name":"TabPFN-3","model_type":"embedding","parameter_count":null,"context_window":null,"claimed_developer":"Prior Labs","confirmed_developer":"Prior Labs","developer_confirmed_at":"2026-05-12T00:00:00+00:00","developer_source":"Confirmed via https://huggingface.co/Prior-Labs/tabpfn_3","first_appeared_at":"2025-01-21T00:00:00+00:00","first_appeared_on":"reddit_machinelearning","status":"confirmed","availability":null,"deployment_status":"not_deployed","availability_scope":null,"gate":3,"eu_ai_act":"GPAI","regulatory_status":{"eu_ai_act":"GPAI"},"open_source":true,"weights_available":true,"license":"TABPFN-3.0 License v1.0 (non-commercial)","provenance_chain":[],"parent_model":null,"model_family":null,"superseded_by":null,"supersedes":[],"recommended_replacement":null,"product_links":null},"events":[{"event_type":"first_appearance","event_date":"2025-01-21T00:00:00+00:00","title":"First recorded in AI news","detail":"TabPFN-3, a pre-trained tabular foundation model, was released as the next iteration of the TabPFN series. The model can now handle up to 1M rows on a single H100 GPU, delivering 10x-1000x faster inference than previous versions. It introduces new capabilities including a Thinking Mode for test-time compute, support for up to 160 classes, calibrated quantile regression, and achieves a 93% win rate over classical ML on TabArena benchmarks. The model is available through three deployment paths: API, enterprise licensing, and open-source weights.","source_url":"https://reddit.com/r/MachineLearning/comments/1tb3fh5/tabpfn3_just_released_a_pretrained_tabular/","source_name":"reddit_machinelearning"}],"delta":{"days_unattributed":476,"claimed_developer":"Prior Labs","confirmed_developer":"Prior Labs","match":true}}