Ascend 0-Day Support for Ling-3.0-flash! New Programming Framework Debuts

Ant Bailing has open sourced the new generation native hybrid reasoning model Ling-3.0-flash, with Ascend completing 0Day adaptation simultaneously and debuting the new operator programming framework CANN PyPTO.

It is understood that the model has 124B total parameters, activating only 5.1B during inference, with computational overhead just 8.1% of the previous flagship Ring-2.6-1T. It alternately stacks KDA (Incremental Attention) and MLA (Multi-head Latent Attention) at a 5:1 ratio, supporting a 256K context window, expandable to 1M.

Meanwhile, the KDA operator in the model is the core structure. Its computation fuses multiple stages of logic including matrix-vector operations, recursive accumulation, and state updates, placing high demands on on-chip data reuse and multi-core parallel scheduling. Traditional hand-written kernel development struggles to balance both efficiency and performance.

To this end, Ascend launched this new programming framework CANN PyPTO, where developers only need to define computation logic through tensor-level interfaces, and the framework automatically handles sharding scheduling, data movement, and code generation — no need to write low-level code, greatly lowering the barrier to operator development.

More notably, with the Ascend CANNBot agent, operator development is decomposed into 7 stages that advance automatically, with each stage executed collaboratively by specialized sub-agents, requiring no manual handoff throughout, significantly shortening delivery cycles for complex fused operators.

On the deployment front, Ascend completed deep adaptation based on the Ascend version of the open-source inference engine vLLM. Patches have been submitted to the open-source platform, supporting lightweight patch deployment with second-level go-live, fully compatible with A2 and A3 series products.

Previously, Ascend had completed adaptation for Bailing Ling-2.6-flash (104B parameters). Ascend stated that this adaptation once again demonstrates its full-process capability to support cutting-edge large models from release to large-scale deployment.