Ling-3.1-flash InclusionAI shows 256,000 on one page and 262K on another

On 30 Sep 2026, TechNode reported that InclusionAI launched a model, and that report is the first half of Ling-3.1-flash InclusionAI. The model is Ling-3.1-flash. TechNode gives it 560 billion total parameters and about 25 billion active per token. It is aimed at agents, search, office software, and specialist apps. It is designed for up to 1 million tokens, but the two-week free trial is capped at 256,000. A larger window and an open-source release are planned after the trial. The same day, Vercel put the model on AI Gateway and printed a different context figure.

TL;DR

What actually happened

Keep the two pages as two pages when you retell Ling-3.1-flash InclusionAI. TechNode is the launch account: who put the model out, what it is for, how big it is, and how the trial is capped. Vercel is a gateway account: where a developer can call it, until which free date, under which IDs, and with which context number on that gateway.

The size lines agree in substance and not in typography. TechNode says 560 billion total parameters and about 25 billion active per token. Vercel says 560B total and 25B active. This page treats those as the same size claim written in long form and short form. It does not treat the context lines that way, because the context lines are not the same digits.

The aim, in TechNode’s account, is agents, search, office software, and specialist apps. That is a target list, not a scoreboard. No benchmark is in the facts used here, so no benchmark is added. “Aimed at” does not mean “measured on.”

The long-context design and the trial cap are both TechNode. Designed for up to 1 million tokens is the design sentence. The two-week free trial is capped at 256,000 is the access sentence. A larger window is planned after the trial. An open-source release is planned after the trial. Planned is not shipped. This page does not move the open-source release into the present tense.

Vercel’s free window is its own date: free through 13 Oct 2026, on AI Gateway. That is not the same sentence as TechNode’s two-week trial cap. One page is talking about a trial limit of 256,000. The other is talking about gateway access that is free through a calendar date, with 262K context on the gateway. A reader can be inside one of those offers and not the other.

The IDs Vercel prints are inclusionai/ling-3.1-flash and inclusionai/ling-3.1-flash-free. Those strings are how the gateway names the routes. They are not a second parameter count.

Sources: TechNode and Vercel’s changelog.

Where the context numbers split

Write the figures the way the pages wrote them. TechNode’s trial cap is 256,000. Vercel’s gateway context is 262K. This article does not rewrite 256,000 as 256K, and it does not rewrite 262K as a spelled-out integer the changelog did not print. Ling-3.1-flash InclusionAI is partly a story about refusing that cleanup.

The design ceiling on TechNode is up to 1 million tokens. That number is larger than both the trial cap and the gateway figure, and it is a design statement, not the limit a caller hits today on either page. If a recap says “1 million context” and drops the trial cap, it has quoted the ambition and hidden the access rule.

After the trial, TechNode says a larger window is planned. It does not, in these facts, say the later window equals 1 million, or equals 262K, or equals something else. “Larger” is the word. Leave it as larger.

The gateway figure, inside any fair account of Ling-3.1-flash InclusionAI, is specifically “262K context on the gateway.” It is not labeled, in the facts used here, as the trial cap, and it is not labeled as the 1 million design. Putting 262K into TechNode’s mouth, or 256,000 into Vercel’s mouth, is how the two pages get falsely reconciled.

Parameter counts are the part that does match. 560 billion and 560B. About 25 billion active per token, and 25B active. You can say the size claim is shared. You cannot say the context claim is shared. That split is the useful finding.

What neither page is

Neither page, in the facts used here, is a benchmark table. No score, no rival model’s number, and no “wins on” line is available to repeat. Anyone who adds a leaderboard under this name is writing a different article. Ling-3.1-flash InclusionAI does not need a fake chart to be worth reading. The disagreement on context is enough.

Neither page is a Polymarket contract. There is no Yes price on whether the open-source release lands, and this note will not imply one. Planned, after the trial, is the tense TechNode used.

The TechNode URL path contains a company string. This article does not upgrade a path into an extra corporate claim beyond InclusionAI launching the model, which is what TechNode’s account is used for here. If you need a fuller ownership sentence, it is not in the facts this page was given.

Free is also not one offer. TechNode describes a two-week free trial with a 256,000 cap. Vercel describes gateway use free through 13 Oct 2026. Those can overlap in calendar time without being the same product surface. Check which door you are walking through before you quote a limit.

Active parameters are “per token” in TechNode’s wording and “active” in Vercel’s short form. This page does not invent a routing diagram to explain the gap between 560 billion total and about 25 billion active. The ratio is the claim. The mechanism, beyond those counts, is not supplied here.

What to do as a reader (not a trade)

Open TechNode and copy the trial sentence with the design sentence still attached. Designed for up to 1 million tokens. Two-week free trial capped at 256,000. Larger window and open-source release planned after the trial. Then open the Vercel changelog and copy 262K as 262K, plus free through 13 Oct 2026, plus the two IDs.

Do not average 256,000 and 262K. An average context is a number neither desk published. The assignment is to leave each page’s figure as that page stated it. Ling-3.1-flash InclusionAI stays honest when those strings stay different.

There is nothing to buy in this note. A model post is not a token, and a free gateway period is not a price target. No profit language fits, and none is added. If you are choosing a tool, choose it by reading the cap on the surface you will actually call.

When you tell someone the size, use both forms once so they can recognize either page: 560 billion total, about 25 billion active per token, and the gateway’s 560B and 25B. Then refuse to give them one context number. Give them two, with labels.

Other posts on this site are easy to confuse with a model launch because they sit nearby. The DogeOS public testnet is not this model. The Gemini 4 Argon launch is not this model. Cite them only as separate notes, never as context-window evidence.

A finished note on Ling-3.1-flash InclusionAI has five anchors. Date 30 Sep 2026. InclusionAI launched it, per TechNode. Size 560 billion total and about 25 billion active per token. Trial cap 256,000 against a 1 million design, with a larger window and open source planned after. Gateway: free through 13 Oct 2026, 262K context, IDs inclusionai/ling-3.1-flash and inclusionai/ling-3.1-flash-free. No benchmarks. That is the file.

If a later page publishes one official context and retires the other, this Ling-3.1-flash InclusionAI snapshot does not update itself. Until that page exists in your hands, the accurate sentence is the split sentence. Say which desk. Say which figure. Say which limit is a design, which is a trial, and which is the gateway. Then stop.