Ant's 560B open-source Ling-3.1-flash tested: builds a city game over 6+ hours

机器之心 · wechat · 2026-10-17

Ant Group's Ant Ling released and open-sourced Ling-3.1-flash (Sept 30 release, Oct 15 open-source). Jiqizhixin ran three hands-on tests of its long-horizon task abilities.

Model basics

560B total parameters, 25B activated per token, up to 1M context
Hybrid linear attention: 7 KDA layers per GatedMLA layer; 512 routed experts with 8 routed + 1 shared expert active per token
ArtificialAnalysis intelligence index 41, 210 tokens/s output, $0.19/M blended tokens
MXFP4 quantized routed experts + block-FP8 dense layers shrink weights from 534GB to 292GB, cutting deployment from 8 to 4 GPUs

Three tests

AI as editor: surfaced and ranked 24h AI news; its top three picks exactly matched the editorial team's real story list that day.
Long-horizon coding: built a city-building game from scratch — runnable prototype in 26 minutes, 6 hours of visual/system iteration, then targeted fixes after human review. Official demo: rewriting a C image library in Rust in 20 hours, passing 30/30 checks with 8x speedup.
On-demand learning tools: from one prompt, generated an interactive webpage teaching Fermat's principle with a draggable light-path simulator, echoing Karpathy's idea.

Ant's stated route: skip benchmark-chasing, pursue real-world AGI, with industry models for healthcare (LingSante) and finance (LingFin).

Original post →

Ant's 560B open-source Ling-3.1-flash tested: builds a city game over 6+ hours

More from coding & agent