Kling 4.0 Preview Leak: Discord Screenshots, Not a Launch

On September 24, 2026, screenshots from Kling's Dev Mode Discord showed two new rows in the creator-app model picker: VIDEO 4.0 Preview and VIDEO 4.0 Flash Preview, both tagged NEW, sitting above the live VIDEO 3.0 Omni / 3.0 / 3.0 Turbo lineup. The settings panel for Preview showed a duration slider from 3 to 30 seconds, labeled up to 10 keyframes — a doubling of the official Kling 3.0 cap of 15 seconds.

That is a leak, not a launch. Kuaishou has not published Kling 4.0. The newest official generation remains Kling 3.0 (announced February 5, 2026). The API changelog's latest entry as of mid-September covered Virtual Try-On 3.0, not a video 4.0. There is no model card. If you are building a pipeline this week, the honest question is wait versus test the preview — not "which Kling 4.0 SKU should we buy."

This post is explainx.ai's dedicated news write-up of that Discord leak. It compares the leaked UI to documented Kling 3.0, to Seedance 2.5's already-shipped 30-second 4K path, and to Runway's research-preview world models that are also not a clip API. For the evergreen platform map, start with our Sora, Runway, and Kling guide.

TL;DR — questions people type after the screenshot

What actually leaked on September 24

Two kinds of material circulated. The stronger evidence is screenshots of Kling's creator interface: a model picker that still lists the current 3.0 family in the right order, plus a Preview settings panel showing 720p and 1080p, four aspect ratios, a 3–30 second slider, "up to 10 keyframes," and a 3 trials left badge. A separate shot of a Long Video mode with copy about generating up to 120 seconds is part of the same dump.

The weaker evidence is a written Discord summary (circulated by aggregator write-ups, including reAPI's September 25 roundup of the same screenshots). That summary adds reference caps, Flash duration, voice-reference counts, and API-package rumors. The author of that summary repeatedly says preliminary settings and no prices are confirmed. That language is the opposite of a spec sheet.

Treat the screenshots as "Kuaishou is testing something labeled 4.0 inside the app." Treat every number from the written summary as a target that can still move.

What Kuaishou has actually published

None of the official channels that would carry a launch name Kling 4.0 as of this writing.

Kling 3.0 is the model you can cite in a contract: 3–15 second generations, 720p through 4K, first-and-last-frame control, native audio, Element control. Native 4K landed on April 23, 2026, months after the February series launch — a reminder that resolution SKUs often trail the version number.

The gap since the June 17 Turbo/Omni notes is the longest quiet stretch of Kling's 2026 video cadence. Combined with selectable Preview rows, that is a reasonable signal that something is close. Close is still not a date, a price, or a model card.

Leaked 4.0 settings vs official 3.0

Put the two columns next to each other without pretending they have equal status. The left column is published. The right column is Discord UI plus a preliminary write-up.

Two things jump out if you build for clients.

Length is the headline, and it is the least you should freeze in a design doc. Doubling 15 seconds to 30 seconds would match a production unit that Seedance 2.5 already claimed as a single pass. Until Kuaishou writes 30 seconds into the capability map, your editor still has to stitch.

4K is missing from the leaked 4.0 panel. That could be a preview-tier omission, a later add (the 3.0 pattern), or a real regression for v1 of 4.0. Do not promise 4K-from-4.0 in a pitch deck.

How this sits next to Seedance, Runway, and the rest of our corpus

explainx.ai already covers the clip-length race from the shipped side.

Seedance 2.5 is the clean comparison for "30 seconds in one generation." ByteDance announced native 30-second 4K, up to 50 multimodal references, and a copyright/provenance story aimed at agencies — in June, at a named conference, with a platform path (Volcano Engine), not a Discord picker. If your job this month is deliver a 30-second 4K spot, Seedance 2.5 is the model with a public claim. Kling 4.0 Preview is a possible future Kling answer to that claim. See the dedicated Seedance 2.5 post and the Higgsfield unlimited Seedance window if you are actually generating volume this quarter (read the 720p cap before you treat "unlimited" as 4K).

Runway in September 2026 is not competing on "one 30-second file." GWM Worlds 2 is a research preview of an interactive 720p world you steer with text — no public clip API, no independent access, same "do not build production on the screenshot" discipline we are applying to Kling 4.0. Aleph-class editing and Characters remain the product surfaces. If your team is choosing finished ads vs interactive worlds, those are different jobs; a Kling 4.0 leak does not collapse them.

Speed and hosted rank tables still describe Kling 3, not 4. fal's August H3 Max write-up compared Kling 3 Standard latency and Elo against MiniMax and Seedance — useful for 15-second hosted traffic, useless as a 4.0 forecast. See fal H3 Max.

Agentic and editor layers do not get a free upgrade from a Preview badge. ViMax still needs a stable generator ID. Palmier Pro's Kling MCP path and video-use still cut and restitch whatever length the API actually returns. A 10-keyframe Preview is interesting for those tools only after the endpoint exists.

Feature-length stitching remains a human pipeline. Fountain 0's Kling-based Odysseus work in our AI Odyssey coverage is hundreds of short clips, not one 120-second experimental mode. Do not read Long Video as "feature film in a box."

Wait vs test: what video builders should do this week

This is the only decision that matters until Kuaishou publishes.

Wait (default for anything that ships to a client)

Stay on Kling 3.0 if you need: a model ID you can put in a runbook, 4K, native audio, first/last frame, commercial terms you can read, and support when a job fails. Design storyboards as 15-second beats (or shorter). Keep your stitcher — Premiere, video-use, or Palmier — as a first-class step, not a temporary hack you plan to delete on launch day.

Do not:

• Quote "30-second Kling 4.0" in a client SOW.
• Hard-code kling-4-0 (or any guessed slug) in production.
• Assume Flash Preview is "Turbo pricing" because 3.0 Turbo existed.
• Treat "3 trials left" as a public beta program. It is a badge on a screenshot.

Watch Kling's API changelog and release notes. Those two surfaces are how 3.0 Turbo and Omni actually arrived. Discord will be louder; the changelog is what your integration tests should subscribe to.

Test the preview (only if you already have access)

If your Kling creator account actually shows VIDEO 4.0 Preview and you have leftover trials, use them like a lab, not a vendor bake-off you will publish as "we evaluated Kling 4.0."

A useful test protocol:

1. Same prompt, same seed if the UI exposes one, on 3.0 vs Preview at 15 seconds — the overlap both systems claim. You are checking identity and motion, not length.
2. One 30-second Preview take of a shot you already failed to hold together by stitching two 15-second 3.0 clips. You are checking whether identity drift across 30 seconds is better than a cut. One clip is not a benchmark.
3. Keyframe stress: if you can place more than first/last, try a three-pose character turn. Count how many of the 10 slots actually bind versus ignore.
4. Resolution honesty: export and inspect 1080p Preview against 4K 3.0. If Preview looks softer, that is data for "wait," not a Twitter dunk.
5. Throw the files away for client work. Preview terms, watermarks, and future spec changes are unknown. Do not archive Preview as the hero of a brand campaign.

If you do not see Preview in the picker, you are not late. You are on the same side of the wall as everyone else. Do not buy Discord "access" from random sellers.

When "wait" and "test" both fail

If you need 30 seconds this week and Kling 3.0 cannot do it in one request, you already have two honest paths: stitch 3.0, or use a model that already documents 30 seconds — Seedance 2.5 being the one we have covered in depth. Hosted promotions can change the unit economics (Higgsfield's Seedance window capped resolution; read that post before you assume 4K volume). That is product choice, not rumor arbitrage.

What people are asking

Is this Garnet, Tango, or Waltz? The Discord summary floated a possible link to anonymous video models seen in testing and immediately said it cannot be confirmed. explainx.ai is not going to merge unnamed eval models into Kling 4.0. If those names appear on an official card later, we will match them then.

Is Kling 4.0 on iOS? Not as a released product. The leak is creator-interface Preview. Phone-app store listings have not been part of any Kuaishou 4.0 announcement because there is no announcement.

Will Image 4.0 ship in the same drop? The written summary mentions Image 4.0 options including high-res output with native 8K unconfirmed. Same rule: UI rumor, not a card. Still-image SKUs often ship on a different calendar than video.

Does 10 keyframes make Kling a "director" tool? More control points are useful if they actually constrain the sampler. First/last frame on 3.0 already fails on hard pose changes. Ten keyframes could be a real upgrade or ten ignored hints. You cannot know from a slider label.

Should agencies pause Kling spend? No. Pause redesigning the pipeline around 4.0. Keep spending on 3.0 if Kling is already winning your motion tests. Switching mid-campaign to an undocumented Preview is how you get unrepeatable heroes.

Where does this sit versus Veo, Sora, Pika? Those comparisons belong in the 2026 video generation guide, which still describes live products. Adding "Kling 4.0" as a row in a quality table would be inventing a model.

Honest limitations (ours and the leak's)

• We did not run VIDEO 4.0 Preview. This is a news synthesis of Discord screenshots and secondary roundups, not a first-party bake-off.
• No model card means no training data statement, no safety evals, no latency SLO, no regional availability, no content-policy diff versus 3.0.
• No independent quality samples with prompts you can re-run. Viral crops of Preview output, if they appear, are selection-biased.
• Long Video 120s is labeled experimental in the leak itself and excludes 2K/4K. Even if it ships, temporal consistency at two minutes is a different problem from a 15-second product shot.
• Aggregator pages that already advertise a Kling 4.0 model ID are placeholders. reAPI's own September 25 article is explicit that Kuaishou has not opened a 4.0 API; any "Kling 4.0" product page that is not live should be treated as a landing page, not access.

What to watch next (without inventing a date)

Kling's 2026 official video cadence, for context only:

A real launch will look like February and June: named post, release-note bullet, capability-map columns, then third-party hosts. Until that happens, explainx.ai will not rewrite the Sora / Runway / Kling guide as if 4.0 were a production default.

If you are assembling an agentic video loop around whatever generator wins, keep the orchestration layer (ViMax) and the cut layer (Palmier Pro) model-agnostic. Version numbers in Discord are the worst possible coupling.

Related on explainx.ai

• AI video generation: Sora, Runway, Kling complete guide — live-product map this leak does not replace
• Seedance 2.5: 30-second native 4K — the shipped 30-second comparison
• Higgsfield unlimited Seedance 2.5 — volume economics and the 720p cap
• Runway GWM Worlds 2 — another preview that is not a clip API
• fal H3 Max: faster-than-realtime hosted video — Kling 3 latency context, not 4.0
• ViMax agentic video generation
• Palmier Pro: Kling MCP video editor
• video-use: Claude Code video editor
• AI Odyssey film: Kling feature-length stitching

Primary leak roundup (not an official Kuaishou spec): reAPI — Kling 4.0 release date: what's confirmed and what leaked (Sep 25, 2026), which cites Kling's API changelog, release notes, Kuaishou's February 3.0 announcement, and the September 24 Discord screenshots.

Kling 3.0 capabilities and the September 24, 2026 Discord leak details are accurate as of September 29, 2026. VIDEO 4.0 Preview numbers are unofficial and may change or never ship. Do not treat this post as a launch announcement.