88Z AI, 22 April — Ant Group's Bailling large model team has released Ling-2.6-flash, featuring hybrid attention and sparse MoE architecture. The model achieves a reasoning speed of 340 tokens per second on four H20 cards, with token consumption at only one-tenth of similar products, while reaching state-of-the-art performance on multiple agent benchmarks.
At a glance
- Ling-2.6-flash uses hybrid attention and sparse MoE architecture with 7.4B active parameters.
- Reasoning speed reaches 340 tokens/s on four H20 cards, with token consumption at about 1/10 of similar models.
The update
Ant Group's Bailling large model team has officially unveiled Ling-2.6-flash, codenamed Elephant Alpha. The model features a total of 104 billion parameters with 7.4 billion active parameters, utilizing a hybrid attention mechanism combining MLA and Lightning Linear along with a sparse MoE architecture. In terms of performance, Ling-2.6-flash achieves a reasoning speed of 340 tokens per second when running on four H20 GPUs. Its token consumption during evaluation is significantly reduced, amounting to only about one-tenth that of similar models. Furthermore, the model has achieved state-of-the-art results on agent benchmarks such as BFCL-V4 and SWE-bench Verified, demonstrating efficient reasoning capabilities and superior performance.
Sources & editorial note
This article is an AI-assisted synthesis by 88Z AI based on the public sources below. Product claims and availability are attributed to their publishers. It does not represent independent testing or first-hand reporting.
- 百灵大模型Source published
Source dates above refer to the original announcements; the date under the headline is this article’s publication on 88Z AI.