Ant Group’s Ling-3.0-flash ships with 124B parameters and only 5.1B active
智东西 · wechat · 2026-07-29
Ant Group’s Ling-3.0-flash pairs 124B parameters with 5.1B active parameters
The article says Ant Group’s Bailing team has launched Ling-3.0-flash, a native hybrid reasoning model that is open for temporary free API use and will be released as open source after the free window.
- 124B total parameters, 5.1B active parameters.
- Reported to outperform the earlier Ring-2.6-1T on 11 of 12 benchmarks.
- Claims top or near-top results on coding, agent, instruction-following, long-context, and tool-use evaluations.
- The model supports 256K context and is positioned for low-latency agent workloads.
- The piece also describes real-world demos: pair-programming a 3D billiards simulator, multi-agent research workflows, Office MCP automation, and design-page generation.
- A separate infrastructure section explains the architecture: alternating KDA and MLA layers, 1/64 expert activation, and distributed caching to cut TTFT by 60%–80% in long-input scenarios.
More from coding & agent
- Engineer Tests Claude for CAD: Turns Messy PPT Specs into DXF Files — burhop · 2026-07-29
- Teams can give coding agents a skills.md file to make them follow repo-specific workflows — Pavan_Belagatti · 2026-07-29
- MCP 2026-07-28 spec goes stateless to scale agent servers — lee_stott · 2026-07-29
- GitHub Copilot app adds stacked PR management inside the agent flow — lee_stott · 2026-07-29
- MCP shifts to a stateless protocol to ease large-scale AI agent deployments — lee_stott · 2026-07-29
- OpenMed Launches Open-Source Medical AI Framework with Local Kimi Integration — MaziyarPanahi · 2026-07-29