Ant Ling's 124B Model Claims Parity With 1T Flagship

Ant Ling released Ling-3.0-flash on July 23, a 124-billion-parameter model that activates just 5.1 billion parameters per token, per Runtime Wire.

The company says the smaller model matches or beats its own 1-trillion-parameter flagship on most benchmarks. It runs on roughly one-eighth the total parameters and one-twelfth the active params per token. If true, that's a real win for anyone paying to serve these models.

The architecture interleaves K-directed attention with multi-layer aggregation layers at a 5:1 ratio and activates just 1/64 of its experts per token. Context runs to 256K natively.

Ling-3.0-flash is live on OpenRouter, free through August 3. Pricing after that date isn't disclosed.

The benchmark numbers come from a founder's X thread, not an independent lab. Ant Ling is comparing a new model to its own older, bigger one and calling it a win.

That makes the free window a test, not a promotion. Independent testers have until August 3 to check the parity claim against the flagship. If nobody reproduces it by then, the pricing silence afterward is the tell.