Atria Dawn Preview

Fri, 18 Sep 2026 00:00:00 +0000

Shanghai AI Laboratory released Atria Dawn Preview on September 11, 2026, an open-weight preview built on a 744-billion-parameter mixture-of-experts foundation and aimed squarely at agentic research and engineering work rather than chat. The design goal is a full loop (problem analysis, tool use, code implementation, running experiments, reading the results, and recovering from failures) rather than answering a question and stopping.

The lab shipped it under an MIT license alongside a 143-author paper studying how the model performed during its own development: across 769 recorded tasks from 56 participants, roughly a third were rated by the humans involved as infeasible without AI assistance.

Ling-3.0-Flash-Fin

Fri, 18 Sep 2026 00:00:00 +0000

Ant Group’s InclusionAI team released Ling-3.0-Flash-Fin on August 28, 2026, continuing the training of the base Ling-3.0-Flash model on financial data developed with CICC and other institutions. It keeps the same 124 billion parameter, 5.1 billion active mixture-of-experts design and 256K context window, but adds source-grounded financial search, multi-document reasoning across regulatory filings and earnings reports, and spreadsheet-native valuation work.

Ant published BF16 weights on Hugging Face under an MIT license and opened a one-month free API trial, aiming the model at tasks like reading annual reports, building DCF models, and cross-referencing filings: the kind of grunt work that eats analyst hours at banks and research desks.

Mit-License on Best of AI

Atria Dawn Preview

Ling-3.0-Flash-Fin