inclusionAI/Ling-3.1-flash model card

Fetched 2026-10-18. Weights repo created 2026-10-13 (HF API createdAt 2026-10-13T11:16:41Z), last modified 2026-10-15, license MIT in the card front matter. Card: hybrid-linear MoE, 560B total / 25B activated, 512 routed experts with 8 activated plus 1 shared expert, 63 KDA + 9 gated-MLA layers, hidden 4096, expert intermediate 1280. The card's README names Terminal-Bench 4.0, SWE-bench Pro and FrontierSWE but prints no numeric table; the numbers live in the HF eval-results widget, sourced from an open PR (discussions/6, commit 49e0e59a, file .eval_results/Ling-3.1-flash.yaml, dated 2026-10-13, source name 'Ling-3.1-flash model card') that is rendered on the model page but not yet merged to main. That YAML reports SWE-bench Pro 68.74 (ScaleAI/SWE-bench_Pro, note 'Run with the LingCode harness'), SWE-bench Multilingual resolved 88.5, Terminal-Bench 2.1 81.18, Terminal-Bench 4.0 46.03 (LingCode harness) and DeepSWE 59.7. All vendor claims, verified=false on the widget.