Ant Ling releases Ling-3.1-flash: writes code continuously for 17 hours, compiler passes 178/182 tests.

动察 Beating AI News Flash: Ant Ling (InclusionAI) has released a brand-new Flash model, Ling-3.1-flash. The model has approximately 560B total parameters, activates only about 25B per token, and can be extended up to a 1M context. Compared with Ling-3.0-flash's 124B total parameters and 5.1B activated parameters, this generation's total scale has expanded by about 4.5 times, and the activated parameters per token are nearly 5 times higher.

Ling also highlighted long-duration coding capability this time. According to the official statement, Ling-3.1-flash spent about 17 consecutive hours writing a Lua-to-x86-64 ELF compiler from scratch, with 178 of 182 tests passing, a success rate of 97.8%. In another task lasting about 20 hours, it ported a C image library to Rust, ultimately achieving an 8.015x speedup, with all 30 correctness checks passing.

The model is already available to try through Novita AI and Vercel AI Gateway. Free access lasts for two weeks, and currently only about a 256K context is available; the full 1M has not yet launched. Ling said that after it shifts to a paid service, the 1M context will be opened, and it plans to open-source the model at the same time.