Ling-3.1-flash: The 560B MoE Ant Group Serves but Has Not Opened

The short answer to what Ling-3.1-flash is: a hybrid-reasoning mixture-of-experts language model from Ant Group's Ant Ling line, around 560 billion total parameters with roughly 25 billion active on each token, a window the vendor caps at 1 million tokens but serves today at 256,000, text in and text out, and no downloadable weights. The model was announced on 2026-09-30 and it is live in Ant's own product and on Ant's own API. That date is where this page starts and where it stops: it is the fact that fixes the model's age, not a launch this article is reporting. What a reader searching for Ling-3.1-flash needs is the state of the thing on the day they read it, and on the day this was written the state of the weights was still "not published".

Three facts do most of the work. The weights are not out, and Ant's own announcement says when it intends them to be. The context figure that circulates — 1M — is a ceiling, not what the endpoint gives you, which is 256K. And the specification is knowable while the price after the promotional window is not. Everything below is sourced, and where a number is Ant's own rather than a third party's, it says so.

One note on where the state of this model lives. Ant's announcement of 2026-09-30 is the document that carries the architecture and the open-weight condition, and it is not something the vendor updates. The pages that do get maintained — the Ling model pages and the pricing page on Ant's developer site, each stamped "Last updated on October 8, 2026" — carry the parameters, the context numbers and the price. Where those two disagree, or where the announcement says more than the documentation does, this page says which one it is quoting.

What Ling-3.1-flash is, and who publishes it

Ling-3.1-flash belongs to the Ling family — the general-purpose language line of Ant Ling, the model programme Ant Group runs and publishes through its inclusionAI lab (the operating entity named on the developer site is Ant Intelligence (Hang Zhou) Technology Co., Ltd.). The family splits three ways: Ling for general language work, Ring for deep reasoning, and Ming for omni-modal understanding and generation. Ling-3.1-flash is the flagship of the Ling branch, sitting above Ling-3.0-flash, which stays in the catalogue as the cheaper, faster option.

Ant positions it narrowly and consistently: long-horizon real-world tasks, office work that runs from a vague request to a finished deliverable, software engineering that spans hours, and agent loops that call tools in sequence and correct themselves. The vendor's own summary of the pitch is that it keeps about 560B of capacity while paying for about 25B per step, so world knowledge and reasoning depth stay while the per-token compute does not. It is the successor to Ling-3.0-flash in the same slot, not a new tier.

Are the weights released? No — and Ant has named the condition

This is the most contested fact on the model and the answer is unambiguous today: the weights are not published. There is no Ling-3.1-flash repository on Hugging Face under the inclusionAI organisation, no model card, no licence text, and no checkpoint file; a search of the organisation's public repositories returns the Ling-3.0 generation and earlier, and no 3.1 entry. ModelScope, where Ant mirrors its releases, has no Ling-3.1-flash model either. Nor is there a deployment recipe for it: the vendor's own Ling Cookbook repository on GitHub, last pushed 2026-10-12, the day before this page, carries local-deploy and quantisation recipes for Ling-3.0-flash, Ling-3.0-flash-VL, Ling-3.0-flash-Fin and Ling-3.0-tiny, and nothing for Ling-3.1-flash. The independent evaluator Artificial Analysis files the model as proprietary, and its FAQ answers "No" to whether the model is open source.

What Ant has done is attach a condition to the opening, and it is more specific than "soon". The announcement states that the free period serves a 256K context, and that after the free period ends and the model converts to paid service, the plan is to open the 1M window and to open-source the model at the same time. So the sequence the vendor describes is: trial at 256K, then paid service at up to 1M, with the weights following at that switch. That is a plan, not a date, and no date is published. Treat any "Ling-3.1-flash is open weights" claim as wrong until a repository exists; treat the plan as the vendor's stated intention, not an event.

Size and architecture: 560B total, 25B active, seven linear layers per global one

Ant publishes the architecture at a level of detail that makes the numbers checkable against each other rather than merely quoted:

• Total parameters — about 560B, vendor-stated, unchanged across the documentation.

• Active parameters — about 25B per token, a sparsity ratio near 1 in 22.

• Expert routing — 512 routed experts, of which each token selects 8, plus one shared expert that every token uses. That is a noticeably wider expert pool than the previous generation used against a much smaller body. The expert count, the 8-of-512 selection and the 7:1 attention ratio come from the announcement rather than the model page; the model page states the parameters and the context numbers.

• Attention — a hybrid stack continuing Ling-3.0-flash's linear-attention design, with the linear share increased: 7 KDA (Kimi Delta Attention) layers to every 1 Gated MLA layer, so that one layer in eight keeps a global attention path. The KDA path is the linear-time one, aimed at long sequences.

• Predecessor for scale — Ling-3.0-flash runs 124B total with 5.1B active over a 42-layer hybrid stack, so Ling-3.1-flash is roughly four and a half times the body with roughly five times the per-token activation.

Worth saying plainly, because Ant says it too: none of this converts directly into deployment cost. Throughput, latency and memory depend on the hardware, the numeric precision, the batch size, expert communication, the context length and the output length. Architecture explains why the model can exist at this size; it does not tell you what a request will cost you.

Context window and maximum output

The window is where the most common misreading lives. Ant's documentation states it as two numbers in one line: the window caps at 1M tokens, and the official service currently exposes 256K. The announcement gives the reason in the vendor's own words — the free trial runs two weeks, the service length during it is 256K, and once the trial ends and the model moves to paid service the plan is to open the 1M window and open-source the model at the same time. The docs and the announcement agree on the served figure, and both are Ant's. The product surface is behind a sign-in, so Ling Studio's own maximum-length control is not something this page could confirm today; if you need the exact cap before you size a pipeline, read it off the studio once you are signed in rather than taking a second-hand number for it.

Artificial Analysis lists the context window as 1M. That is the declared ceiling, not what a request currently gets, and it is the number that travels furthest and is wrong most often in practice. If you are sizing a pipeline, size it to 256K and treat 1M as a roadmap item with no published date.

Maximum output is a figure Ant does not publish per model. What the vendor does publish is the request surface: the Anthropic-compatible endpoint's max_tokens defaults to 32,000, the OpenAI-compatible reference lists no per-model output cap, and non-streaming calls time out at 90 seconds. If your job needs a long single generation, the vendor's own guidance is to stream rather than to raise a cap it does not document.

What it takes in and what it gives back

Text in, text out. Ling-3.1-flash is not a multimodal model, and it is worth being explicit because the family contains one that is: Ling-3.0-flash-VL is the vision model, natively accepting image, text and video input through the same platform, and the vendor's multimodal documentation is written for that model. Ling-3.1-flash appears in the vendor's Ling specification table without any input modality beyond text, and the independent evaluator's page for it records text input and text output only, and files it among reasoning models rather than vision ones. If you need to send an image to a Ling model today, the vendor's own selection guide points you at Ling-3.0-flash-VL, not at this one.

Vendor claims versus independent evaluation

Two kinds of number exist for Ling-3.1-flash and they are not interchangeable.

Ant's own announcement is titled as a model upgrade for real-world tasks, and it is careful about what its numbers are. The named evaluation areas are GDPVal-AA V2.1 for office work, HealthBench Professional for medical communication, FrontierSWE for engineering, Terminal-Bench 4.0 for terminal operation, plus automation, safety analysis, codebase question answering and financial-agent tasks — and the per-benchmark figures sit in chart images rather than a text table. As of today Ant has not published those individual scores in a form anyone can cite line by line, and this page will not invent a table for them. What Ant does state in text are three long-horizon engineering results, all vendor-reported, all on the vendor's own task definitions, and all labelled as the comparison figures "listed in the documentation" rather than as a matched head-to-head: a Lua-to-x86-64-ELF compiler built from scratch in about 17 hours of continuous execution, with 178 of 182 independent tests passing, about 97.8%, against 99.0% for Claude Fable 5 and 95.9% for GLM-5.3-Flash on the same task as the vendor lists them; a Pyright 1.1.400 optimisation pass of roughly 10 hours reaching a 1.432× speedup with all 2,279 benchmark and regression tests passing, against Claude Fable 5 at 1.502× and GLM-5.3-Flash at 1.290×; and a Rust reimplementation of a C image library taking about 20 hours, passing all 30 checks at 8.015×, against GLM-5.3-Flash at 1.396× and Claude Fable 5 at 27.310×. Those three tasks are not benchmarked head-to-head at matched effort and harness — the opposing figures are the vendor's own citations, not runs it controlled — so read them as the vendor's report of its own task suite, not as a like-for-like ranking. Ant also reports that the model found and patched a stack-buffer-overflow in the open-source OceanBase database's calibration parser by working the problem autonomously. Every one of those is a first-party claim on a first-party task; none has a third-party reproduction, and the vendor itself flags the third result as showing headroom left in extreme performance optimisation. On the search demand side, "ling 3.1 flash benchmark" is one of the smaller questions people ask about this model, and the honest one-sentence answer is the one above: Ant's card exists, its figures are vendor-reported, and the independent number you can actually hold the model to is the next paragraph.

The second kind of number is independently measured. Artificial Analysis, which runs its own harness on hardware it controls, scores Ling-3.1-flash at 41 on its Intelligence Index v4.3.2, a composite of ten evaluations, and ranks it fifth of 175 models on that index. The predecessor scores 20 on the same index. The same page records 212.1 output tokens per second and 1.87 seconds to first token, 220 million output tokens generated across the index — high against a median of 95M — and a running cost of $0.30 per million input tokens and $0.90 per million output with an 80% cache discount, about $1.00 per index task. Those are the evaluator's own measurements against the vendor's API, not Ant's figures, and they are the only independent measurements of this model that exist today.

On price, the same split applies, and the currency matters. Ant's pricing page — the maintained one, stamped "Last updated on October 8, 2026" rather than the 2026-10-13 date this page was written — lists Ling-3.1-flash at ¥0.00 for input, output and cache reads, and states that the model is free for a limited-time two weeks, after which it switches to standard paid billing with the rates to be announced. The same page publishes a USD table in which the row reads $0.00 across all three columns. That is the whole of the free-tier answer: free is real, it is promotional, it is time-boxed, and the number that follows it has not been published. The independent evaluator's recorded rate of $0.30 and $0.90 is the closest thing to a post-trial rate card in public, and Ant has not confirmed it.

How to get a response from Ling-3.1-flash today

There are two first-party routes, both Ant's own. The Ling Studio at chat.ant-ling.com is the product surface, and the vendor's announcement also says the model is inside an internal Ant desktop agent used for office tasks. The Open API at api.ant-ling.com is the developer route and ships in two shapes: an OpenAI-compatible chat/completions endpoint and an Anthropic-compatible messages endpoint, so an existing SDK works with a base-URL change. Two caveats come straight from the documentation, which is unusual enough to be worth copying exactly: the probability and diversity controls, temperature and top_p, default to 1 in both reference pages, so a request that has learned to set a low temperature elsewhere will behave differently here; and the thinking control documented for toggling reasoning is documented as Ling-3.0-flash only, which means the request surface these pages publish has not been updated for Ling-3.1-flash. Access needs an Alipay account bound and the Ling service activated — the FAQ also lists "Ling model service not activated" as a known error with its own fix — billing is pay-as-you-go per token, the published rate limit is 2 requests per second globally with 429s beyond it, and non-streaming calls time out at 90 seconds. The vendor also lists the model at the same promotional zero price on third-party platforms; those are not first-party surfaces and their terms are their own.

One verification worth doing before you build on the endpoint: as of today the vendor's published API reference does not enumerate Ling-3.1-flash in its model list, and the platform changelog carries no entry for it — the newest entries there are Ling-3.0-flash-VL in September and Ling-3.0-tiny in August — even though the pricing page carries a rate for the model and the announcement offers its API trial. Confirm the model identifier against your own account before you wire it into anything, rather than assuming the documentation has caught up.

Ling-3.1-flash is not on OrcaRouter's catalogue — a check against our live model list on 2026-10-13 returns no Ling or inclusionAI entry — so this page will not tell you it is one call away. What is one call away are the two models it is most often weighed against: DeepSeek V4.1 Flash, which our list prices at $0.15 per million input and $0.60 per million output, and GLM 5.3 Flash at $0.15 and $0.50, both at list price with zero markup and automatic failover across providers, so a rate change from either vendor lands the same day without a code change on your side. When Ant opens Ling-3.1-flash or a host adds it, that is the shape it would arrive in.

What to watch, and what not to assume

Three things settle this page's open questions and all three are Ant's to publish: the weights, with a licence and a model card if the announcement's plan holds; the 1M window actually being served rather than declared; and a post-trial rate card. Until at least the first two land, the accurate description of Ling-3.1-flash is a strong, fast, closed, text-only model with a genuine independent score of 41 and a genuinely promising agentic profile, available first-party and free for a window that closes without warning.

Who should act on it now: anyone whose work is agentic, terminal-shaped or office-shaped and who wants two weeks of free inference on their own prompts to find out whether a 560B MoE at 212 tokens per second changes what they build. Who should wait: anyone planning to self-host it, anyone who needs a licence before procurement will look at it, anyone whose pipeline needs the million-token window in production, and anyone who needs a cost model that outlives the promotional period. For all four of those, the honest answer today is that the decision cannot be made yet, and the thing that would unblock it is a Hugging Face repository, not another benchmark chart.