1. 30 seconds is the entry ticket: why these four
Division of labor first: in August this site already ran two video-model comparisons. One is the AI video generation review 2026-08, which lined up five closed-source cloud models: Sora in its shutdown countdown, Veo 3.1, Kling 3.0, Jimeng, and Runway. The other is Seedance 2.5 vs MiniMax H3, which unpacked the two routes shipped on the same day, July 31. This piece does not repeat that work. It takes a new angle: by October 2026, "single-clip 30-second native direct output" has turned from a few players' signature feature into a standard ticket for front-runners, so the selection question has shifted from "can it generate one complete finished clip in a single pass" to "whose 30 seconds fits your job."
Following four quadrants--cloud flagship, active benchmark, cloud value, and open self-hosting--this review picks four models that represent the current landscape: Kling 4.0 (Kuaishou, the cloud-flagship heir presumptive, not yet officially launched), Seedance 2.5 (ByteDance, the already-live active benchmark), Wan 3.0 (Alibaba Cloud, the cloud value pick, officially live since August 24), and LTX-2.5 (Lightricks, the only open-weight, self-deployable route among the four). The reference date throughout is 2026-10-03.
Boundaries first: every comparison below is compiled from official announcements, official documentation, and media reports, not from my own stress tests. With no stress-test data, this piece refuses to arbitrate "who is number one"; it maps each quadrant's range and hands down per-need verdicts at the end. All prices carry snapshot calibers and defer to official pages. Two frequently asked models are absent: Sora (conflicting sources on its status; this site's August review holds the baseline) and the current Veo version (not verified first-hand).
2. Spec comparison: four models side by side
| Dimension | Kling 4.0 (Kuaishou) | Seedance 2.5 (ByteDance) | Wan 3.0 (Alibaba Cloud) | LTX-2.5 (Lightricks) |
|---|---|---|---|---|
| Launch status | Launching in October; specs below are officially announced | Live since 2026-07-31 | Officially live since 2026-08-24 | Weights released (custom community license) |
| Clip length | Up to 30s, extendable to 2 min via continuation | 30s native direct output, multi-turn extension | Specifiable 2 to 30s | Pipeline-dependent, default 24fps output |
| Reference materials | Up to 15 items (10 images, 5 videos, 7 subjects) | Up to 50 items (30 images, 10 videos, 10 audio) | Direct document upload (doc/xls/ppt/pdf/md) | Keyframe, IC-LoRA, and other conditioned pipelines |
| Editing | Precise video editing, multi-keyframe | Local editing, timestamp precision under 1s | Natural-language plus timestamp shot planning | Retake regeneration, DubIt re-voicing |
| Open source / self-host | No | No | No | Yes, 66GiB weights self-deployable |
| Price caliber | Not announced | Per official release materials | API from 0.3 RMB/sec (August snapshot) | Weights free; commercial revenue threshold USD 10M |
| Best for | Teams waiting for big-scene, pro-grade specs | Creators who need long-narrative clips now | Budget-sensitive batch production | Teams with GPUs needing data privacy |
The row to watch is "launch status": only two of the four are fully usable today; one is counting down and one is DIY. That shapes how to read the next four sections--judge shipped models on performance, unlaunched ones on announced specs, and the open-weight one on its thresholds.
3. Cloud flagship heir: Kling 4.0, specs maxed but doors not open yet
Kling 4.0 is the only one of the four not yet officially live. Per Kling's official announcements and media reports (Shanghai Securities News, Cailian Press, and others in late September), Kling 4.0 was announced on September 28 with an October launch date; the same day, the lightweight Kling 4.0 Flash opened a small-scale trial for Black-Gold annual-card members. Read this section accordingly: lay the specs out, and hold judgment until hands-on.
The officially announced specs are substantial: 4K and 1080p output in 10-bit HDR; single-generation clips up to 30 seconds; multi-turn continuation stretching to 2 minutes; a prompt input ceiling raised to 8,000 tokens. Multimodal references go up to 15 items, combinable in one run as 10 images, 5 videos, and 7 subjects. Language coverage spans Chinese, English, Japanese, Korean, Spanish, Portuguese, German, French, Hindi and more, including dialect accents, with lip matching, text rendering, and multi-keyframe support. The creation page was upgraded in tandem: a unified input box, grid and list layouts, timeline-based continuous creation, and a new canvas Agent experience.
Mapped against shipped products, Kling 4.0 is betting on "professional-grade specs": 10-bit HDR and 2-minute continuation target film production and ad-delivery pipelines, while the 8,000-token prompt ceiling leaves room for feeding a full multi-shot script in one pass. The competitive backdrop is stated plainly in media coverage: it faces the pressure created by Seedance 2.5's launch in July; Citi kept its Neutral rating on Kuaishou, watching exactly two indicators--whether Kling can pull away Seedance's user base, and commercial pricing competitiveness. Pricing is unannounced and this piece will not guess it. A Shanghai Securities News reporter noted lukewarm early external impressions; likewise, no judgment on an unlaunched product here. Follow the launch in this site's hotspot piece Kling 4.0 official launch, and use the getting-started guide How to use Kling AI when it lands.
4. Active benchmark: Seedance 2.5, the default for long-narrative clips
If you only look at long-narrative capability that works "right now," Seedance 2.5 is the steadiest row on this list. Per its official release materials: formally released on July 31, 2026, live the same day in Doubao Pro and Jimeng, and recently opened to enterprise users through Volcano Engine. Three core items: single-clip 30-second native direct output with multi-turn extension; up to 50 multimodal reference materials per run, combinable as 30 images, 10 videos, and 10 audio clips, with white-model and green-screen references supported; and local editing of backgrounds, products, and characters with timestamp precision under one second. On language, it synchronizes lip shape and speech rate across more than ten languages; Jimeng's ultra-long generation mode reaches 3 minutes, and Maya and Blender plugins connect it to professional pipelines.
Deployment evidence speaks louder than parameters: XCMG uses it for industrial training SOP videos, XPeng plugged it into its design platform, and Weifen Zhifei generates embodied-AI training data with it. These cases point one way--30-second native direct output is already in production environments, not just keynote demos. Note that Seedance 2.5 is not open source; it ships as closed-source API and products. Open weights are a different line--do not confuse it with LTX-2.5 here or with MiniMax H3 in the site's head-to-head piece.
5. Cloud value pick: Wan 3.0, a 30-second quote sheet at 0.3 RMB per second
Wan 3.0 has the clearest value story of the four, with facts following this site's established record and Alibaba Cloud Bailian's official documentation: officially live on August 24, 2026, single clips up to 30 seconds, with any integer duration from 2 to 30 seconds specifiable. Its differentiated entry point is on the input side: it first supports direct upload of five office formats--doc, xls, ppt, pdf, and md--so a product-introduction deck can become a 30-second promo video directly, a genuine workflow simplification for enterprise users. Text-to-video, image-to-video (first frame or first-and-last frames), and reference-to-video paths are all open, with multi-shot storyboards described in natural language plus timestamps.
Price is its sharpest edge (late-August snapshot caliber, defer to official pages): the Bailian API bills per second, at 0.3 RMB/sec for 480P, 0.6 RMB/sec for 720P, and 1.2 RMB/sec for 1080P; a 30-second 720P clip lists at 18 RMB. The launch-period 30-percent-off promotion ran through September 23 and has ended as of this writing, so estimate at list price. The boundary also needs stating: per Alibaba Cloud's developer documentation, it does not support Function Calling, web search, or batch inference, and calls are submitted asynchronously--it is a per-second generation engine, not an Agent, and workflow orchestration stays on your layer. Details are in this site's Wan 3.0 official launch piece.
6. Open self-hosting: LTX-2.5, the only one you can pull onto your own GPU
Only LTX-2.5 ships open weights, which is why it occupies its own quadrant. Capability, per the GitHub repository README: LTX-2 is the first DiT-architecture joint audio-video foundation model, generating picture and synchronized audio in one pass; the recommended LTX-2.5 distilled transformer is 22B parameters (bf16 safetensors), paired with a custom text encoder, video and audio VAEs, and spatial and temporal upsamplers, with the full weight set at roughly 66GiB. Inference efficiency is its signature: the DistilledPipeline uses only 8 preset sigmas, and the distilled 8-step path is enough to ship output; the default output is 1024x1536 at 24fps, with a 4K tier at 3840x2176. Twelve pipelines cover production needs: keyframe interpolation, IC-LoRA conditioned video generation, Retake for regenerating local time segments, DubIt for re-voicing while preserving lip shape and speaker identity, native EXR and HDR, and more. For tuning, the official ltx-trainer package supports LoRA, full fine-tuning, and IC-LoRA, and ComfyUI has an official plugin.
Thresholds and licensing must be stated plainly: it is not OSI-certified open source but a custom LTX Community License. Individual non-commercial use is free, and commercial entities' test-and-evaluation work is exempt; but any commercial use by an entity with annual revenue of USD 10 million or more requires purchasing a Commercial Use Agreement from Lightricks, and derivatives--including fine-tuned weights and LoRA adapters--must ship under the same terms. Self-hosting also means supplying your own GPU resources sized for 66GiB of weights. What you get back is what the other three cannot offer: data stays inside your network, offline inference, and fine-tuning for your own business. Details are in this site's LTX-2 open source piece.
7. Per-need verdicts
With no stress-test data, this piece will not crown a "number one"; verdicts hang on your requirement.
Want a ready cloud flagship: pick Seedance 2.5. It is the only one of the four already live with 30-second long narrative, 50-item referencing, and sub-second local editing all running, and its deployment cases are verifiable in production. If you need footage today, it is the shortest path.
Want pro-grade specs and big scenes: wait for Kling 4.0, but do not wait idle. 10-bit HDR, 2-minute continuation, and 8,000-token prompts genuinely target film and ad pipelines, but it is unlaunched, unpriced, and unverified beyond official announcements. Put it on the watch list and validate with real projects once it lands.
Want value and batch production: pick Wan 3.0. Per-second billing from 0.3 RMB plus direct document upload make it the friendliest quote sheet for enterprise batch scenarios among the four, at the cost of no Agent capability and self-built workflow orchestration.
Want data privacy and self-hosting: pick LTX-2.5. The only open-weight option, with 8-step distillation lowering the inference bar and ltx-trainer for fine-tuning--fit for teams with GPU resources and hard data-compliance requirements. Remember the license premise: free for individual non-commercial use, but commercial entities above USD 10 million annual revenue must buy the agreement.
8. FAQ
Q1: Can Kling 4.0 be used right now? A: Not fully. Per Kling's official announcements and media reports, it is set to launch in October; only the lightweight 4.0 Flash is in a small-scale trial for Black-Gold annual-card members. All Kling 4.0 specs in this piece are officially announced figures, with real-world performance pending launch.
Q2: Which of the four is truly open source? A: Only LTX-2.5 ships open weights, and note it is a custom community license, not OSI-certified open source: free for individual non-commercial use, while entities with USD 10 million or more in annual revenue must purchase a commercial agreement. Kling 4.0, Seedance 2.5, and Wan 3.0 are all closed-source cloud deliveries.
Q3: Does "30 seconds" mean the same thing across the four? A: Not exactly. Seedance 2.5 and Kling 4.0 both cap a single clip at 30 seconds--the former usable today, the latter pending launch; Wan 3.0 lets you specify 2 to 30 seconds; LTX-2.5's clip length depends on the pipeline and VRAM. On top of that, Kling 4.0 offers continuation stretching to 2 minutes and Seedance 2.5 has a Jimeng-side ultra-long mode reaching 3 minutes--neither is a direct counterpart to "a single 30-second clip."
Q4: Which should a budget-limited team pick? A: For batch production, look at Wan 3.0 first: API from 0.3 RMB/sec (480P, August snapshot caliber), a 30-second 720P clip at 18 RMB list. Teams with GPUs and a willingness to self-operate can add LTX-2.5 to the cost model: weights are free, but VRAM, operations, and licensing thresholds belong in the math.
Q5: Are these prices final? A: No. All prices here are snapshot calibers: Wan 3.0 reflects a late-August Bailian page snapshot, Kling 4.0's pricing is unannounced, and LTX-2.5's threshold figure comes from the license text. Always defer to official pages before signing anything.
Interaction
Which member of the 30-second club are you using? Waiting on Kling 4.0's HDR specs, already delivering with Seedance 2.5, running batches on Wan 3.0, or pulling LTX-2.5 onto your own GPU? Share your selection reasoning and pitfalls in the comments; frequent questions will roll into future updates. If this helped, pass it to a colleague mid-selection.