Ling-3.2-flash vs Ling-3.1-flash: Same 560B Skeleton, One New Sense, No New Weights

Ant Ling's own deprecation page makes the Ling-3.2-flash decision harder than the launch post does. Three older models — Ling-2.6-1T, Ring-2.6-1T and Ling-2.6-flash — are being retired on 15 November 2026, and the recommended replacement for all three is Ling-3.1-flash. So on the day Ling-3.2-flash arrived, its predecessor had just been designated the official landing zone for every team migrating off the 2.6 generation. Those teams now have to choose between the model Ant told them to move to and the model Ant announced the same week. The two share a parameter count, share a context window, share an API surface — and differ on exactly one axis that matters, plus one that matters more: Ling-3.2-flash sees, Ling-3.1-flash reads, and Ling-3.1-flash is the one you can download.

That is the comparison worth making, and it is not the one the launch coverage made. Ling-3.2-flash wins on capability. Ling-3.1-flash wins on everything a production team needs in order to commit: a published price, MIT-licensed weights, and an independent score you can check — even if that score has since been withdrawn from the leaderboard it lived on.

What is genuinely identical

Start with the parts Ant did not change, because they determine how much work a migration actually is.

• Size — both are 560B total parameters with 25B active per token. Ant's model-selection table lists the same two figures on both rows.

• Context — both carry a 1M-token window. Ling-3.0-flash, one tier down, is capped at 256K, so 1M is now the flagship baseline rather than a headline.

• Request surface — both are served at api.ant-ling.com/v1 through an OpenAI-compatible chat/completions endpoint and an Anthropic-compatible messages endpoint. Both accept the thinking object with enabled or disabled; Ant's reference lists 3.0, 3.1 and 3.2 series alike.

• Lineage — same hybrid-linear MoE family, same KDA-to-gated-MLA attention recipe Ant introduced at 3.0 and refined at 3.1, same 512-routed-expert topology as the open 3.1 checkpoint.

• Rate limits — the same global ceiling of two queries per second with 429s above it, and the same 90-second synchronous timeout with streaming recommended. A new model does not buy you more throughput.

So a migration from Ling-3.1-flash to Ling-3.2-flash is a model-string change plus whatever you do differently with images. There is no new SDK, no new endpoint, no new auth.

What actually changed: the model learned to look

Ling-3.1-flash is text in, text out. That is not a limitation Ant hid — the vendor's own selection guide positions it for office work, long-horizon software development and complex agent tasks, and its 2026 launch material was full of engineering benchmarks: a Lua-to-x86-64-ELF compiler port, a Pyright conformance pass, a Rust reimplementation, an OceanBase parser patch. It was a model for people who write code for a living.

Ling-3.2-flash is repositioned around four visual capabilities: complex visual understanding, multimodal long-horizon agents, 3D understanding and interaction, and medical multimodal work. The API reference is explicit that a Ling-3.2-flash request "natively accepts text, images, and videos," using image_url blocks for stills and video_url blocks for video. The demos Ant published show what that is for: reading headless renders of material spheres and comparing roughness and normal maps to refine a script; driving Blender through UI actions; converting engineering drawings into 3D web models; turning research PDFs with equations into working code; estimating portions and calories from meal photos.

Ant's three release benchmarks are all visual, and all vendor-reported. OmniDocBench1.5 at 92.21 against 90.56 for Opus 5 at maximum effort; ZEROBench(Sub) at 43.71 against 39.52; Vision2Web at 85.86 against 80.20. The claim they support is narrow and specific: on document parsing, 3D-scene reasoning and vision-to-web implementation, Ant believes the new model beats a frontier closed model. Not one of the three says anything about text reasoning, and Ant published no text benchmark at all — which is itself a signal, given that Ling-3.1-flash's launch was almost entirely text benchmarks.

If your workload is text, the honest reading is that Ling-3.2-flash's advantage over Ling-3.1-flash is unmeasured. If your workload involves looking at something, it is the whole point.

The price question is where the two diverge hardest

Ling-3.1-flash has a published rate card, currently discounted. Ant's pricing page lists it at ¥2.00 per million input tokens and ¥6.00 per million output, marked down 75% to ¥0.50 and ¥1.50, with cache reads at ¥0.40 falling to ¥0.10. The USD column Ant publishes for third-party platforms shows $0.30 and $0.90 at list, discounted to $0.075 and $0.225.

Ling-3.2-flash has no rate card at all. The pricing page lists Ling-3.1-flash, Ling-3.0-flash-VL, Ling-3.0-flash and two Ming image models; Ling-3.2-flash appears only in a promotional banner: "free for the first 2 weeks after launch," with original prices resuming afterwards and a suggestion to watch your account balance. No input rate, no output rate, no cache rate, and — more consequentially for a multimodal model — no published treatment of image or video tokens.

For anyone planning capacity, that is the deciding asymmetry. A model you cannot price is a model you cannot put in a budget, and a vision model whose image-token accounting is undocumented is a model whose cost you cannot even estimate from its text rate. The free window closes around 19 November, five days before the Ant deprecation wave lands on the 15th. Teams migrating off Ling-2.6 will hit both dates inside the same fortnight.

Weights: the predecessor is open, the successor is not

This is the part that reverses the intuitive ordering. The older model is the one with downloadable weights.

The Hugging Face repository inclusionAI/Ling-3.1-flash is MIT-licensed and carries 572,120,680,768 bytes of safetensors — roughly 572 GB — last modified 3 November 2026. Its sibling Ling-3.1-flash-1019, the post-training revision Ant pushed in late October, is also MIT and holds 54 FP8 shards at about 573.8 GB. Both are the full checkpoint, not adapters, and both require trust_remote_code because of the custom BailingMoeV3ForCausalLM modelling code.

Ling-3.2-flash has no repository. The inclusionAI organisation's recent listing runs Ling-3.1-flash-1019, Ling-3.1-flash, a base checkpoint, the ZwZ2 vision models and Ming audio models — no 3.2 entry. Under Ant's announcement the most-liked reply is a user wishing it were open source, and Ant has not answered.

So the two models trade places depending on what you need. If you want to serve a 560B MoE on your own hardware, evaluate it offline, fine-tune it, or simply not depend on a vendor's uptime, Ling-3.1-flash is available under terms that permit all of it, and Ling-3.2-flash is not. If you want the better multimodal model and are content to rent it, the ordering flips.

The independent scores, such as they are

Artificial Analysis rebuilt its Intelligence Index to v5.0.1 in late October, and the rebuild left the two Ling models in different states. Read today, the Ling 3.1 Flash page carries no Intelligence Index figure at all — the tile reads N/A, alongside N/A for time per index task, cost per index task and output tokens. The Ling 3.1 Flash 1019 page, by contrast, carries a live score of 30 on v5.0.1, with $3.19 weighted cost per index task and 240M output tokens generated.

That 30 is not comparable to any pre-rebuild Ling number. Under the previous rubric Ling-3.1-flash-1019 read 44; the v5 rubric is harder, weights 60% of the composite toward private and held-out sets, and it moved most models down. The 44-to-30 shift is re-scoring, not regression.

Ling-3.2-flash has no Artificial Analysis page at all. There is no third-party number, no speed measurement, no cost-per-task estimate. The only figures in existence are Ant's three visual benchmarks.

How to decide

The decision rule that follows from all of the above is unusually clean, because the two models are not really substitutes.

Stay on Ling-3.1-flash if your workload is text and your timeline is production. You get a published discounted rate, MIT weights you can self-host, an independent score of 30 on the current rubric for the 1019 build, and Ant's own recommendation as the replacement for three retiring models. Nothing about Ling-3.2-flash improves on any of that for text work, except an unmeasured possibility.

Move to Ling-3.2-flash if the work involves images, video, documents with layout, 3D or screenshots. That is the capability Ling-3.1-flash simply does not have, and it is the only reason to accept an unpublished price and closed weights. Run your evaluation during the free window — the integration is a model-string change on an SDK you already have — and treat anything you build on it before 19 November as a pilot rather than a commitment.

Run both if you have mixed traffic. Route the visual work to 3.2 and the text work to 3.1, which is what the two-model split is actually for, and note that Ant's rate limit is global per account rather than per model — the ceiling does not double because you are calling two models.

Where a router fits, and where it does not

Being precise about our own position, since it is easy to imply more: neither Ling-3.2-flash nor Ling-3.1-flash is on the OrcaRouter catalogue. We re-read the live model list on 6 November and it returns 215 identifiers with no Ling, inclusionAI, Ring or ERNIE entry — the only "ling" matches are Kling video models. So nothing in this comparison is one call away today.

What a router does buy you here is the exit path rather than the entry. Ant's rate limit is two queries per second per account; a global limiter on a 560B flagship is exactly the condition where failover earns its keep, and the practical way to run Ling-3.2-flash in an agent loop is to put a second provider behind it rather than to queue behind a 429. On our side that second provider is a model with a published rate on the same key: Qwen3.8-Flash at $0.15 per million input and $0.47 per million output for vision work, or DeepSeek V4.1 Flash at $0.15 and $0.60 and GLM-5.3-Flash at $0.15 and $0.50 for text. Prices are provider list passed through at zero markup, so a vendor cut lands the same day instead of after a repricing cycle. The routing DSL composes several models into one call, which is how the split above becomes a single endpoint instead of two integrations.

What would settle it

Three publications would turn this from a judgement call into a spec comparison. A rate card for Ling-3.2-flash, including how image and video tokens are counted. A Ling-3.2-flash repository under the same MIT terms as its predecessor, which would make the older model's main advantage disappear overnight. And a text-only benchmark table — one reasoning suite, one coding suite — that would tell us whether the 3.2 generation improved the thing 3.1 was built for, or only added a sense to it.

Until then, the ordering is: Ling-3.2-flash is the better model, and Ling-3.1-flash is the better bet. Ant shipped the two of them five weeks apart — Ling-3.1-flash on 30 September 2026, Ling-3.2-flash on 5 November — and only one of those two models has a price on it.