Ant Ling launches Ling-3.2-flash: 560B/25B native multimodal MoE (image/text/video), 1M context, API-first

Why it matters

【Key points】

• Ant Group’s Ant Ling (inclusionAI) on 2026-11-05 (~19:40 UTC) launched Ling-3.2-flash via @AntLingAGI and docs: latest native multimodal MoE — ~560B total / ~25B active, up to 1M context, native image/text/video.
• Focus: complex vision, multimodal long-horizon agents, 3D interaction, medical multimodal.
• Live on Ant Ling’s OpenAI-compatible API (`Ling-3.2-flash`); OpenRouter example `inclusionai/ling-3.2-flash`; Computrix trial.
• No HF open-weight page found this round (unlike Ling-3.1-flash weights e/878).
• Vendor graphics claim wins vs Opus 5 (max) on OmniDocBench1.5, ZEROBench(Sub), Vision2Web — vendor-reported only.

【Our read】

1. Background: 3.1 (e/878) pushed the same MoE into long-horizon agents/coding. 3.2 wires native multimodal into that activation budget so image/video are reasoning inputs, with 3D and medical called out. Ant shipped docs+API+demos as a product drop.
2. Comparison: Google productized Omni 1.2 Flash for scalable video (e/1090); OpenAI’s image path is Images 2.6 (e/1088); Qwen-VL/GLM-V/Kimi vision keep pressure. Ling-3.2 bets large MoE ratio + 1M + agent/3D/medical in one API. Missing: open weights and independent evals; Opus 5 numbers are vendor release metrics.
3. Impact: Near term, trial via Ant Ling/OpenRouter/Computrix. Watch: weights following 3.1’s open path, external vision/agent evals, and whether medical multimodal becomes a compliance product.

【Verifiable】

• Docs (2026.11 Ling 3.2 / Ling-3.2-flash)
• API lists `Ling-3.2-flash` with image_url/video_url
• Official X2026-11-05T19:40:27.000Z
• OpenRouter guide

• On-site: e/878 Ling-3.1-flash weights; e/977 Ming-3.0.
• HF `inclusionAI/Ling-3.2-flash` not found.
• Not Ling-3.0-flash-VL.