Ling 3.1 Flash: Ant Group’s 560B Model Specs, Pricing, and Test

👁 121 reads

Ant Group, the parent company of Alipay, has released Ling 3.1 Flash: a mixture-of-experts model with roughly 560 billion total parameters, activating around 25 billion per token. For roughly two weeks following its debut, you can test it for $0 on OpenRouter. We looked into the pricing, global availability, and ran the model through 3 quick multilingual test tasks: it scored 11 out of 12 points, including flawless invoice math.

What Ling 3.1 Flash Can Do

Ant Group announced Ling 3.1 Flash on September 30, 2026, and the model appeared on OpenRouter on October 2. It is the first “flash” tier model in this lineup exceeding 500 billion parameters. With 25 billion active parameters out of 560 billion, inference is far cheaper computationally than the gross size suggests. The model is a hybrid architecture - it can return direct answers immediately or engage an internal reasoning phase first (reasoning is enabled by default on OpenRouter).

Ant Group positions this model primarily for autonomous agent workflows, search retrieval, and enterprise office automation.

Benchmark scores published by Ant Group (these are vendor-reported figures, not independent audits):

• 1,673 Elo in GDPVal-AA v2.1 (office and enterprise task benchmarks),
• 75.16 in FrontierSWE (software engineering and coding),
• 65.35 in HealthBench Professional (medical domain benchmarks).

Aggregator Opper cites an Artificial Analysis Intelligence Index score of 41.1, which we have not verified at the original source.

Ling 3.1 Flash Pricing

Pricing figures are taken directly from the OpenRouter catalog in US dollars.

The free promotional window is slated to last roughly two weeks, roughly through mid-October, though no official expiration date is listed. Some downstream API intermediaries have already begun charging their own markup, so free access depends on the specific provider. Our benchmark run cost $0.

Hands-On Performance Test

To evaluate its non-English linguistic and reasoning abilities, we ran the model through 3 targeted tasks in Polish: writing a cohesive narrative incorporating a regional idiom and city declension, calculating a commercial utility bill with a 23% VAT rate, and proofreading a paragraph packed with 6 common grammatical and spelling blunders. This provides an initial snapshot rather than our exhaustive test suite from the multilingual AI model arena. For a direct baseline, we submitted the identical prompts to Mistral Large 4.

Idiom and Grammatical Inflection

We asked for a paragraph about a business trying to cut heating expenses, requiring the Polish idiom equivalent of “coming away empty-handed” (wyjść jak Zabłocki na mydle - literally “to come out like Zabłocki on soap“) and the proper locative form of the city Bielsko-Biała. Excerpt from the generated text:

Inicjatywa, która miała obniżyć koszty, zakończyła się fiaskiem - sprawa po prostu wyszła jak Zabłocki na mydle. Takie eksperymenty z budżetem w Bielsku-Białej zwykle kończą się boleśnie dla portfeli. (“An initiative meant to cut costs ended in failure - the whole thing simply went down the drain. Such budget experiments in Bielsko-Biała usually end painfully for wallets.“)

Inflections were spot-on, and the phrasing sounded natural. The idiom was framed slightly loosely (attributing the outcome to an abstract matter rather than an individual), but the core meaning of financial loss was preserved accurately.

Invoice and Tax Math

We provided a two-month utility bill totaling 312.40 gross with a standard 23% VAT rate. Ling 3.1 Flash calculated the net sum of 253.98, the VAT amount of 58.42, and the monthly split of 156.20 - exact down to the last decimal point and formatted precisely as requested.

Proofreading

We submitted a corrupted paragraph: „Wczoraj poszłem do sklepu (…) Niewiem czy zdąże przed zamknięciem bo sklep zamykają o osiemnastej. Pani w kasie powiedziała że wziąść paragon trzeba zawsze.” (“Yesterday I went [incorrect masculine form] to the store (…) I don’t know [spelled as one word] if I’ll make it [missing nasal e] before closing because they close the shop at six. The cashier said you always have to take [archaic non-standard infinitive] a receipt.”) Ling corrected the compound negation, nasal endings, infinitive form, and missing commas, but it missed the incorrect past-tense verb poszłem (Mistral Large 4 missed this exact error as well). It also unnecessarily altered the plural verb for store hours from plural to singular.

What the Numbers Do Not Show

The free API endpoint suffers from severe traffic congestion. During our initial test run, all three queries failed immediately with HTTP 429 rate-limit errors from upstream host Novita. We had to retry three times before securing a full set of completed responses.

How to Access the Model

• OpenRouter: fully functional worldwide under the model ID inclusionai/ling-3.1-flash, routed through Novita as of October 7, 2026. You only need a standard OpenRouter account and API key.
• Vercel AI Gateway: platform aggregators indicate the model became available there on launch day, also relying on Novita routing. We have not verified this route independently.
• Ant Group Web Interface: there is currently no standalone consumer chat portal in English or European languages hosting this specific model.
• Downloadable Model Weights: not yet released. inclusionAI stated plans to open-source model weights after the initial evaluation period. Because predecessor Ling 3.0 Flash was released on Hugging Face under an MIT license, an open release is likely, though no target date has been set.

We could not find an official data retention policy for this preview endpoint, so you should avoid submitting confidential or proprietary data.

Who Is Ling 3.1 Flash Best Suited For?

• Agent and workflow developers: a high-capacity model supporting function calling at $0, ideal for prototyping before final pricing is announced.
• Multilingual drafting: dependable enough for quick drafting, though human review remains necessary for detailed copyediting.
• Not ready for critical production: constrained by strict preview rate limits, lack of finalized SLA pricing, and absent offline weights.

Further Reading

• OpenRouter - how to access multiple LLMs through a single unified API
• Chinese AI Models Evaluated - our comparative multilingual test across 5 leading models
• Lesser-Known AI Models - exploring MiMo, Qwen3.8 Max, GLM-5.3, and Muse Spark
• AI Model Arena - our comprehensive 13-task benchmark

Google · Your sourcesQuantslant higher in your Google - one clickAdd to preferred sources →

Sources

• OpenRouter: Ling 3.1 Flash - pricing, context limits, and provider data
• Ant Ling (@AntLingAGI) on X - launch announcement and official benchmarks
• TechNode, Sept 30, 2026 - architecture specs and preview timeline
• OrcaRouter - status of downloadable weights
• Hugging Face: Ling 3.0 Flash - licensing and previous open weights

Tested on October 7, 2026. Early availability and pricing can shift quickly after release - we will update this guide once Ant Group publishes permanent rates or opens weight downloads.

Kinetyra