Ling-2.6-flash Open-Sourced

Ling-2.6-flash has been officially open-sourced. It is an instruct model with 104B total parameters and 7.4B active parameters.

• It replaces the existing GQA with a 1:7 MLA + Lightning Linear hybrid architecture and combines it with sparse MoE to boost inference efficiency.
• On a 4× H20-3e setup (TP=4, Batch Size = 32), it recorded up to 340 tok/s, and prefill and decode throughput were reported to improve by up to about 4x.
• In the overall Artificial Analysis evaluation, it reportedly maintained competitive performance while using only 15M tokens.
• It reportedly showed strengths in agent benchmarks such as BFCL-V4, TAU2-bench, SWE-bench Verified, Claw-Eval, PinchBench, as well as in general knowledge, math reasoning, instruction following, and long-context understanding.
• It provides a recommended SGLang configuration along with vLLM support, and for the MTP (Multi-Token Prediction) version, a separate patch branch was noted due to a bug in official SGLang.
• Remaining limitations noted include tool hallucination in complex scenarios, Chinese-English switching, and the need for improvement in following complex instructions.