Overview
Ling-3.2-flash is a native multimodal model from Ant Ling, released on November 5, 2026, with 560B total parameters, 25B active, a 1M context window and native image, text and video input. This generation did not buy capability with parameter count: total and active sizes sit in the same bracket as the previous Ling-3.1-flash. What changed is that vision moved from a bolted-on module into the architecture itself, so images and video now feed reasoning directly.
Ant Ling groups the capability into four areas: complex visual understanding for deep reasoning over images and video, multimodal long-horizon agents that hold a goal across multi-step tasks using vision and text together, 3D understanding and interaction for parsing scenes and spatial relations, and medical multimodal work covering imaging interpretation and report reading.
In release evaluations three benchmarks come out ahead of Claude Opus 5 (max): OmniDocBench1.5 at 92.21 against 90.56, ZEROBench(Sub) at 43.71 against 39.52, and Vision2Web at 85.86 against 80.20. On the DesignArena image-to-frontend arena it enters the top ten with an Elo of 1,238, alongside models from Meta, DeepSeek and Anthropic. Those three figures are vendor-reported.
Access is direct: the API is live on the Ant Bailian platform, compatible with both the OpenAI and Anthropic protocols, and listed on OpenRouter as inclusionai/ling-3.2-flash, free for the first two weeks. Pricing is $0.30 per million input tokens and $0.90 per million output tokens.
Key Features
- 560B total, 25B active: A MoE whose total-to-active ratio matches the previous Ling-3.1-flash, so per-step compute follows 25B while knowledge capacity sits at 560B.
- Native multimodal input: Vision and language are fused at the architecture level, with image, text and video all accepted as reasoning input through the image_url and video_url fields in the API.
- 1M context window: Long documents, long video and multi-turn tool use can be handled in one pass, which is the foundation for the multimodal long-horizon agent work.
- Complex visual understanding: OmniDocBench1.5 scores 92.21, aimed at structured visual content such as document layouts, tables and formulas.
- 3D understanding and interaction: ZEROBench(Sub) reaches 43.71, covering scene parsing and spatial relations, one of the newly named capabilities in this generation.
- Medical multimodal: Covers medical imaging interpretation and report reading, positioned to assist professional analysis rather than answer patient questions.
- Both protocols supported: Compatible with OpenAI and Anthropic calling conventions, so existing clients and SDKs switch over with a new base_url and model name.
Use Cases
- Document and report parsing: 92.21 on OmniDocBench1.5 points at layout, table and formula fidelity, suited to batch structuring of invoices, research reports and contracts.
- Image to frontend: turn a design file or screenshot into a working page, where an Elo of 1,238 on the DesignArena image-to-frontend arena is competitive.
- Video understanding: video is native input, so summarizing, searching and questioning long footage needs no frame extraction step first.
- Multimodal long-horizon agents: sustained multi-step operation inside a 1M window, holding the goal with vision and text together.
- 3D and spatial tasks: scene understanding and spatial relation judgment where three-dimensional perception is required.
- Medical imaging assistance: joint reading of images and reports inside existing clinical workflows.
Pros
- Native multimodal rather than a stitched pipeline, with image and video as first-class inputs that need no second alignment pass.
- 560B total with 25B active leaves room between capacity and calling cost.
- A 1M context paired with multimodal input means long video and long documents go in whole.
- Pricing at $0.30 input and $0.90 output per million tokens sits at the low end for this performance tier.
- Compatibility with both OpenAI and Anthropic protocols keeps switching costs low.
- The first two weeks are free, enough to run real samples before committing to spend.
- The platform is directly reachable from mainland China without extra network workarounds.
Pricing
Billed per token: $0.30 per million input tokens and $0.90 per million output tokens, free for the first two weeks after launch. The API is open on the Ant Bailian platform and compatible with both OpenAI and Anthropic protocols; on OpenRouter the model id is inclusionai/ling-3.2-flash, and third-party platforms also offer trials. This round ships as an API only, with open weights not part of the release.
Summary
Ling-3.2-flash does not add parameters so much as wire native multimodality into an existing activation budget, which turns image and video into reasoning input rather than a bolt-on. Three vendor benchmarks clear Opus 5 (max), the DesignArena image-to-frontend arena puts it in the top ten, and pricing sits at $0.30 input and $0.90 output per million tokens. For document parsing, image to frontend or long video understanding, the free two weeks are enough to run a real sample set.
Version History
- Ling-3.2-flash ships native multimodality into a 1M context (2026-11-05): Ant Ling released the native multimodal model Ling-3.2-flash on November 5, 2026, with 560B total parameters, 25B active, a 1M context window and native image, text and video input, strengthening complex visual understanding, multimodal long-horizon agents, 3D understanding and interaction, and medical multimodal work. Three vendor benchmarks come out ahead of Claude Opus 5 (max): OmniDocBench1.5 at 92.21 against 90.56, ZEROBench(Sub) at 43.71 against 39.52, and Vision2Web at 85.86 against 80.20. It enters the top ten of the DesignArena image-to-frontend arena with an Elo of 1,238. The API is live on the Ant Bailian platform, compatible with OpenAI and Anthropic protocols, listed on OpenRouter as inclusionai/ling-3.2-flash, free for the first two weeks and priced at $0.30 input and $0.90 output per million tokens.