Chinese Models Desk
Chinese Models Desk

Ant Group's Ling-3.2-flash Adds Eyes to a 560B Model, Gives the API Away for Two Weeks, and Keeps the Weights Home

Ant Group's InclusionAI has launched Ling-3.2-flash, a native multimodal 560B mixture-of-experts model with 25B active parameters and a 1M-token context window. The API is free for two weeks, and the open-weight question is the one that matters most.

John EditorNov 13, 2026 9m read
ShareWhatsAppXFacebook

Ant Group has quietly done something that most Western labs would have announced with a keynote: it took a 560-billion-parameter language model that developers already liked, taught it to see, and put it behind a free API. Ling-3.2-flash, released on November 5, 2026 by Ant's AGI arm InclusionAI, is a native multimodal mixture-of-experts (MoE) model that accepts image, text and video input, activates only 25B parameters per token, and carries a 1M-token context window.

The commercial story is as interesting as the technical one. The model is free through the API for the first two weeks, speaks both the OpenAI and Anthropic protocols, and arrives without open weights. For a lab whose recent reputation rests on MIT-licensed releases, that last detail is the one to watch.

What Ant Actually Shipped

According to the Ling model page on Ant's developer portal↗, Ling-3.2-flash is "the latest-generation native multimodal model in the Ling series." The headline specifications, all taken from Ant's own documentation, are:

  • 560B total parameters with 25B activated per token, the same size class as the text-only Ling-3.1-flash.
  • A 1M-token context window, which Ant says supports long-document and long-video analysis.
  • Native input of image, text and video, fused "at the architectural level" rather than bolted on through an external vision module.
  • Four stated focus areas: complex visual understanding, multimodal long-horizon agents, 3D understanding and interaction, and medical multimodal work.

The Ant Ling changelog↗ dates the launch to November 5, 2026 (platform version v2.0.6). It describes the model as available through the chat interface and through API calls "compatible with both OpenAI and Anthropic protocols," and says it is free for the first two weeks after launch. Two weeks from November 5 lands on November 19, although Ant has not posted a precise end date.

The architecture, as far as it is public

Ant has not published a technical report for Ling-3.2-flash, so the architecture details below come from its siblings, and should be read as context rather than as confirmed specs for the new model. The Ling-3.1-flash model card on Hugging Face↗ lists the same 560B/25B scale as a hybrid-linear MoE: 63 KDA layers and 9 Gated MLA layers (a 7:1 ratio), 512 routed experts, one shared expert and 8 activated experts per token. Ant's own roadmap table notes that the Ling 3.0 generation introduced KDA (Kimi Delta Attention) to improve memory update and information retention over long sequences.

That lineage explains how a 1M-token window is affordable. Linear-attention layers keep memory growth in check, while a small number of full-attention layers preserve precise retrieval. Whether Ling-3.2-flash uses exactly the same layer mix is something Ant has not said.

"Retains 560B model capacity while keeping per-step activation at around 25B, balancing model capability with inference efficiency." — Ant Ling developer documentation

The Benchmarks: Selective, Vendor-Reported, and Interesting

Ant's launch post on X claimed the model "outscores Opus 5 (max) in release evaluations," a claim summarized by llms.blog's launch write-up↗. Three results were highlighted:

  • OmniDocBench 1.5: 92.21 for Ling-3.2-flash versus 90.56 for Claude Opus 5 (max), a document-parsing test that matters directly for enterprise paperwork.
  • ZEROBench (Sub): 43.71 versus 39.52, a hard visual-reasoning subset.
  • Vision2Web: 85.86 versus 80.20, a test of turning visual input into working front-end code.

Those are the wins. The same source points out that the chart attached to Ant's post had 12 test panels and five rival models, and that Opus 5 (max) scored higher on several others, including Humanity's Last Exam-MM (42.69 vs 30.12), AndroidWorld (90.09 vs 85.63), BlenderBench (79.44 vs 67.77) and the full ZEROBench set (26.00 vs 18.00). The chart contained no text-only or coding tests.

So the honest reading is nuanced. Ling-3.2-flash looks strongest at document understanding and design-to-code, and weaker than the top Western model at deeper agentic and expert-level multimodal reasoning. These are Ant's own numbers. There is no technical report, and the same outlet found no independent reproduction at the time of writing.

What outside observers have added

llms.blog also relays a report from Chinese outlet IT Home that the model entered the top ten of the DesignArena image-to-frontend arena with an Elo of 1238. That is third-party framing of a vendor claim, and it is consistent with the Vision2Web result, but it is not an independent evaluation of the whole model.

Treat every number above as a launch-week claim. Ant's own chart is more candid than its headline: the model wins where documents and web pages are involved, and trails where long expert reasoning is.

Where It Sits in the Ling Family

Ling-3.2-flash is the newest entry in a fast cadence. Ant's evolution table on the model page runs from Ling 1.0 in March 2025, through the trillion-parameter Ling 2.0 in October 2025, to Ling 3.0 in July 2026, Ling 3.1 in October 2026, and Ling 3.2 in November 2026. That is a major generation roughly every quarter, and a point release every month or so.

The Startrise release tracker↗ lists five recent InclusionAI releases, and four of them shipped with downloadable weights. The pattern is revealing:

  • Ling-3.1-flash (September 30 announcement) was a roughly 560B hybrid-reasoning MoE offered through the API first, with MIT-licensed weights following on October 13.
  • Ling-3.1-flash-1019 upgraded that checkpoint, and the tracker reports MIT weights on Hugging Face on October 27, with an Artificial Analysis Intelligence Index score of 44 versus 41 for the earlier version.
  • Ling-3.0-flash-VL was Ant's first native multimodal model: 124B total, 5.5B active, 256K context, open-sourced under MIT on Hugging Face↗.
  • Ling-3.2-flash is the first to combine the 560B/25B scale with vision, and so far it is API and chat only.

If the 3.1 pattern repeats, an API-first launch followed by weights a couple of weeks later would be consistent with history. But Ant has not promised it for 3.2, and nobody should treat a pattern as a commitment.

How to Try It

Ant's multimodal understanding guide↗ documents the practical limits, and they are tighter than the 1M-token headline suggests:

  • Images: JPEG and PNG, passed as Base64 only, with up to 40 images per request and a 32 MiB cap on the request body.
  • Video: MP4, MOV or WMV, no longer than 30 seconds, one video per request, sampled at a fixed 2 frames per second and capped at 32 frames.
  • Protocols: the OpenAI-compatible Chat Completions API takes `image_url` and `video_url` content blocks, and the Anthropic-compatible Messages API also accepts text, images and video for Ling-3.2-flash.
  • Endpoint: the API reference↗ documents the chat completions route, and the model identifier in the multimodal guide is `Ling-3.2-flash`.

Note the gap between the claim and the interface. A model marketed for "long video analysis" currently accepts a single clip of at most 30 seconds through the documented API. The 1M-token window is real for text and documents, but video workloads will hit the frame cap long before they hit the token cap. That may change, but today it is the documented behavior.

The Business Angle: Free Now, Priced Later

For a commercial reader, the two-week free window is the most important detail. It is a classic land-grab: remove the cost of experimentation, let agent builders and front-end tool vendors wire the model into their pipelines through the Anthropic-compatible endpoint, and then switch the meter on.

Pricing is the open question. The llms.blog write-up says IT Home reported US$0.30 per million input tokens and US$0.90 per million output tokens, but also notes that Ant's own pricing page listed no Ling-3.2-flash rate as of November 8 and said original prices would resume after the promotion. Until Ant publishes a rate card, the IT Home figures should be treated as a media report, not a price list.

If the reported numbers hold, the model would undercut Western frontier multimodal models by a wide margin while claiming parity on selected tasks. That is the same playbook DeepSeek and Alibaba have run for text models, now applied to vision.

Who benefits

  • Document-heavy enterprises in finance, insurance and logistics, where an OmniDocBench lead translates into fewer parsing errors on invoices, contracts and scans.
  • Front-end tooling companies, since design-to-code is exactly where Vision2Web and DesignArena results point.
  • Healthcare-adjacent builders, with a caveat: Ant's documentation describes the medical capability as assisting "professional medical analysis" and publishes no medical test results, so this is not a clinical-grade claim.
  • Agent developers, who can use the long context and the Anthropic-compatible protocol to drop the model into existing coding-agent harnesses.

The Competitive Picture

Ling-3.2-flash enters a crowded week. The llms.blog report notes that Baidu released ERNIE 5.5 Flash Base weights the same day, November 5. The broader competitive question is whether Chinese labs are converging on a standard formula: a very large sparse model, a small active-parameter footprint, a million-token window, aggressive API pricing, and open weights as a distribution strategy.

Ant's twist is vertical focus. Rather than chase one general leaderboard, it picked four applied areas (visual understanding, agents, 3D and medical) that map to products Ant and its partners already run, including the consumer app named in its Ling-3.0-flash-VL documentation. That is a commercially coherent bet. It is also a less testable one, because most of its public evidence comes from benchmarks that Ant selected.

What would change the assessment

Three developments would sharpen the picture over the next few weeks:

  • An open-weight release. Weights under MIT would let independent labs reproduce the benchmark claims and let enterprises self-host, which is the strongest argument Ant has against closed Western rivals.
  • A published technical report and an official price list. Both are missing today, and both are prerequisites for serious enterprise procurement.
  • Independent evaluations on the benchmarks where Opus 5 (max) led, which would show whether the weaker results are a quirk of Ant's chart or a real ceiling.

Bottom Line

Ling-3.2-flash is a credible, well-scoped multimodal release, not a leaderboard-wide breakthrough. It combines scale (560B total, 25B active), reach (1M tokens), and an aggressive distribution tactic (free for two weeks, OpenAI and Anthropic compatibility). Its self-reported strengths in document parsing and design-to-code are plausible given the lineage of the Ling family, and its self-reported weaknesses are visible on Ant's own chart.

The decision for developers is easy for the next two weeks: test it while it is free, especially on document and front-end workloads, and keep your integration portable. The decision for everyone else waits on one question: whether Ant treats Ling-3.2-flash as another open-weight Ling model or as the first Ling model it keeps. The answer will say a lot about how far commercial pressure is shifting the open-source habits of China's AI labs.

#Ant Group#InclusionAI#Ling-3.2-flash#Multimodal#Mixture of Experts

Comments

Open discussion — no account needed. Be respectful.

0/4000
Loading comments…