From the source
Ant Ling released Ling-2.6-flash, an Instruct model with 10.4 billion total parameters and 7.4 billion active parameters.
The model uses a hybrid linear attention architecture (MLA + Lightning Linear in a 1:7 ratio) to achieve inference speeds up to 340 tokens/s on 4-H20 cards and prefill throughput 2.2 times that of Nemotron-3-Super.
It consumed only 15 million tokens in an Artificial Analysis evaluation, roughly one-tenth the tokens used by Nemotron-3-Super, while achieving performance comparable to or state-of-the-art on agent benchmarks such as BFCL-V4, TAU2-bench, SWE-bench Verified, Claw-Eval, and PinchBench.
