AI Briefing
KOSign in

Ant Ling releases Ling-3.2-flash: 560B/25B native multimodal MoE with 1M context

·2026.11.07 14:21

Key point

The 560B-parameter model supports text, image, and video inputs and achieves 92.21 on OmniDocBench 1.5.

Details

  • Ant Ling released Ling-3.2-flash, a native multimodal Mixture-of-Experts (MoE) model with 560B total parameters and 25B activated parameters.
  • The model supports a maximum 1M-token context window and accepts text, image, and video inputs.
  • Evaluation results include 96.80 on MathVision (w/CI), 92.21 on OmniDocBench 1.5, and 85.63 on AndroidWorld.
  • AntBench-Medical performance is reported as 89.52% comprehensiveness and 99.22% correctness.
  • API access is available through both OpenAI-compatible and Anthropic-compatible interfaces.
  • Demonstrated capabilities include PDF-to-code implementation, GUI operation, 3D reconstruction, and multi-agent collaboration.
multimodalmoelong-contextapi-releaseant-lingling-3.2-flash
Read originalShare on Threads

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.

Related

RedditAgnes AI Releases Agnes-3.0-Qwen: 180B MoE Model with 1M Context and Strong Agentic BenchmarksRedditSensenova releases SenseNova-Loop-9B-Preview with 256K context and looped architectureHuggingFace BlogMeta Releases Muse Spark Open: 816B-Parameter Open-Weight Multimodal MoE ModelKakaoKakao Releases Open-Source 3B Multimodal Decision-Making Model 'Kanana-2-v-decide'PyTorchKRThinking Machines Lab Releases Active 17B Inkling 2 Small Tinker and Plans to Open WeightsHacker NewsMemorilla Compresses LLM Memory into 16 Tokens for Better Long-Context Performance