Ant Group’s Ant Ling team launched Ling-3.2-flash on November 5, 2026: a native multimodal Mixture-of-Experts model with 560 billion total parameters, about 25 billion activated per token, a 1-million-token context window, and native image, text and video input. Official docs say the release focuses on complex visual understanding, multimodal long-horizon agents, 3D understanding and interaction, and medical multimodal tasks. Ant reports vendor scores above Claude Opus 5 (max) on three release evaluations, and the API is free for the first two weeks.
Quick Take
- What: Ant Ling released Ling-3.2-flash, a 560B / 25B-active multimodal MoE with 1M context and native image/text/video input.
- Vendor benches: OmniDocBench1.5 92.21 vs Opus 5 max 90.56; ZEROBench(Sub) 43.71 vs 39.52; Vision2Web 85.86 vs 80.20 — all Ant-reported.
- Access: Chat experience plus OpenAI- and Anthropic-compatible APIs; free for two weeks after launch per the changelog.
- Why it matters: Another Chinese frontier multimodal MoE is pitching vision-heavy agent loops (PDF-to-code, photo-to-3D style demos) at API speed, not only chat demos.
What happened
Ant Ling (@AntLingAGI) posted the launch with the 560B / 25B split and the three Opus 5 comparisons. The platform changelog dated November 5 matches those specs and adds the two-week free API window plus dual-protocol compatibility. Model docs at developer.ant-ling.com describe native multimodal fusion without an external vision bolt-on, and position 3.2-flash as the latest native multimodal line after Ling-3.1-flash (text/agent-oriented) and earlier 3.0 VL variants.
Background: Ling’s MoE ladder
Ant Ling’s public timeline runs from domestic-heterogeneous MoE validation in 2025 through trillion-parameter Ling 2.x work and the 2026 Ling 3.x hybrid-linear generations. Ling-3.2-flash keeps the “large skeleton, small activation” pitch: 560B capacity with ~25B active per step, matching the efficiency story Ant used for 3.1-flash while adding stronger native vision/video. Prior open-weight drops in the family (such as Ling-3.0-flash-VL on Hugging Face) do not automatically imply 3.2-flash weights are open; the November 5 materials emphasize API and chat access.
How it is meant to be used
Ant’s demos and docs stress closed-loop visual work: inspect intermediate renders, correct scripts, drive 3D tools, turn engineering drawings into web models, and read research PDFs into code. That is a different product bet than text-only coding agents. Long 1M context is pitched for long documents and long video. Builders should treat the three headline benches as vendor claims until third parties reproduce them, and should confirm pricing after the promotional window ends — secondary posts circulate dollar-per-million figures that are not the focus of Ant’s English model card.
| Spec / claim | Ling-3.2-flash (official / vendor) |
|---|---|
| Total / active parameters | 560B / ~25B |
| Context | 1M tokens |
| Inputs | Image, text, video (native) |
| API | OpenAI + Anthropic compatible; free first 2 weeks |
| OmniDocBench1.5 | 92.21 vs Opus 5 max 90.56 (vendor) |
| ZEROBench(Sub) | 43.71 vs 39.52 (vendor) |
| Vision2Web | 85.86 vs 80.20 (vendor) |
Pros and cons
For users
- Pro: A free two-week window lets teams try multimodal PDF, UI and 3D-ish workflows without a new vendor negotiation.
- Con: Chat/API-only access (no confirmed open weights for 3.2-flash in the launch notes) limits offline and custom fine-tune paths.
For developers
- Pro: OpenAI- and Anthropic-compatible endpoints reduce harness friction for existing agent stacks.
- Con: Vendor-only benches and demo-led marketing mean you must bring your own evals for document parsing and vision-to-code quality.
For companies
- Pro: Another competitive multimodal API option in a market where Claude Opus-class vision pricing is a budget line item.
- Con: Data residency, medical-claim liability and independent security review still sit with the buyer — “medical multimodal” on a model card is not a clinical certification.
What X is saying
Launch amplification focused on the Opus 5 comparisons and the free window.
Introducing Ling-3.2-flash, Ant Ling’s native multimodal model: 560B total / 25B active parameters. Stronger complex visual understanding, long-horizon multimodal agents, 3D understanding & interaction, and medical intelligence.
— Ant Ling (@AntLingAGI) November 5, 2026 · View on X
A 25B-active model is already beating Claude Opus 5 Max on multiple multimodal benchmarks. Ant Ling just introduced Ling-3.2-Flash… OmniDocBench 1.5: 92.21 vs. 90.56
— Emergent AI (@jialiwanai) November 6, 2026 · View on X
Ant Ling just shipped Ling-3.2-flash: 560B MoE / 25B active, native image+text+video, 1M context. Vendor card says it beats Opus 5 (max) on OmniDocBench… Free API for two weeks. No open weights this round.
— Sai Teja (@saiitoshii) November 6, 2026 · View on X
Ant’s Ling-3.2-flash IS OUT. A native multimodal model that beats Claude Opus 5 (max) across multiple benchmarks… 560B total, 25B active per token, 1M context
— 智东西China AI News (@Chinazhidx) November 6, 2026 · View on X
What Grok says
We did not find a dedicated standalone Grok analysis thread of Ling-3.2-flash at the time of writing. Treat Ant’s release evaluations as claims to reproduce, not as settled rankings.
Critical analysis
Ling-3.2-flash is a serious product-shaped release: clear parameter split, dual API protocols, a dated changelog and a focused multimodal thesis. The weak points are familiar. The Opus 5 wins are vendor-reported on a narrow visual set; they say little about text reasoning or terminal agents. “Medical multimodal” should be read as a capability theme, not a regulatory status. And without open weights, the competitive lever is price/latency/quality on Ant’s endpoint after the free fortnight. Independent OmniDocBench and Vision2Web runs will decide whether this is a durable challenge to Opus-class vision or a launch-week chart.
What’s still open
- Will third parties reproduce the three headline scores?
- Post-promo pricing, rate limits and data-processing terms?
- Open-weight plans for 3.2-flash, if any?
- How DesignArena and other live arenas move after the first week of traffic?
About the analysis
Researched and written with AI assistance under the editorial responsibility of the i10X News team. The title image is an AI-generated illustration. Read our AI Use Policy. Spotted an error? Request a correction.
This independent, research-based i10x analysis is based on Ant Ling’s official model docs and November 5, 2026 changelog, the @AntLingAGI launch post, and secondary coverage that restates vendor benchmarks. No Ant/Alibaba company hub is assigned on i10x; filed under Models. More related multimodal coverage lives under i10x news topic pages as published.
i10x Perspective
The multimodal MoE race is consolidating around a simple buyer question: can the model look at an intermediate artifact and correct itself cheaper than your current vision stack? Ant is answering with a 25B-active flash SKU and a two-week free trial. Believe the demos when your own PDFs and UIs survive the loop — not when a vendor slide beats Opus by two points.
