According to CoinMeta, Ant Birdling ( inclusionai ) has released a brand-new flash model called Ling-3.1-flash. This model has approximately 560 billion total parameters, with only about 25 billion of these parameters being activated per token, and it can be scaled up to a maximum of 1 million contexts. Compared to the previous generation's 124 billion total parameters and 5.1 billion activated parameters, this new model has a total scale that is about 4.5 times larger, with each individual token having nearly 5 times as many activated parameters. Officials claim that Ling-3.1-flash was able to write a compiler from scratch in just 17 hours, covering the range from lua to x86-64 elf, passing 178 out of 182 tests with a success rate of 97.8%. In another task that took about 20 hours, it managed to port a C image library to rust, achieving an 8.015-fold acceleration, and all 30 correctness checks were passed. The model is now available for free through novita ai vercel ai gateway, with a free trial period lasting for two weeks. Currently, only about 256 thousand contexts are available; the full 1 million context version has not yet been launched. Birdling stated that after transitioning to a paid service, they will open up 1 million contexts and plan to make the model open-source simultaneously.