88Z AI, 6 November — Ant Bailin has released Ling-3.2-flash, a native multimodal model with 560B total parameters and 25B activated per step, supporting a 1M context window. It features visual feedback self-check and adaptive anomaly recovery, allowing it to correct biases using rendering or interface results and retry after errors or tool failures. Using a hybrid attention mechanism of 7-layer KDA and 1-layer Gated MLA, it achieves an average decoding speed of 280 tokens/s for single users with long contexts under
At a glance
- Ling-3.2-flash features visual feedback self-check and adaptive anomaly recovery for bias correction and automatic retries.
- It uses a hybrid attention mechanism, achieving an average long-context decoding speed of 280 tokens/s under DSpark speculative inference.
The update
Ant Bailin has officially launched Ling-3.2-flash, a native multimodal large model. It features a total parameter count of 560B, with 25B parameters activated per step, and supports a context window of up to 1M. Functionally, the model introduces visual feedback self-check and adaptive anomaly recovery capabilities. This allows the model to use rendering or interface results as observational inputs to correct biases, or to automatically locate issues and retry in the event of errors or tool call failures. Regarding architecture and performance, the model employs a 7:1 hybrid attention mechanism combining 7 layers of KDA with 1 layer of Gated MLA. Supported by DSpark speculative inference technology, it achieves an average single-user decoding speed of 280 tokens/s when processing long contexts of approximately 30K, significantly improving efficiency in long-text processing.
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.