BytePlus September 2026: Seedream 5.0 Flash, Seedance 2.5 Premium and Two New Reasoning Models
Disclosure: this post contains an affiliate link to BytePlus. If you sign up through it we may earn a commission at no extra cost to you. Everything below.
BytePlus had a quiet September in public and a busy one in its docs. There was no blog post, no press release and no launch thread — just three new models appearing on the ModelArk Model releases page and a plugin release on GitHub. If you build on BytePlus, this is the update you missed.
One practical note before the list: the ModelArk docs are rendered entirely in JavaScript, so they are invisible to most scrapers and to the AI tools people use to track model launches. That is why this has not been written up elsewhere. We loaded the pages in a real browser to read them.
Three new models on ModelArk in September
| Model ID | Provider | Type | What it is |
|---|---|---|---|
dola-seedream-5-0-flash-260915 | ByteDance | Image generation | The same capability set as Seedream 5.0 Pro — interactive editing, layer decomposition, native multilingual generation — but faster and cheaper. BytePlus positions it for latency- and cost-sensitive applications. |
deepseek-v4-1-flash-260910 | DeepSeek | Deep reasoning | 552B total parameters, mixture-of-experts, with native multimodal visual understanding. BytePlus describes it as a lightweight flagship that outperforms mainstream flagship models including DeepSeek-V4-Pro on multiple benchmarks. |
glm-5-3-flash-260828 | Zhipu | Deep reasoning | The first native multimodal model in the GLM-5 series. 320B total parameters with 18B active, combining sparse and linear attention. |
Every description above is BytePlus’ own. None of these claims has independent benchmarking that we could find, and the docs do not publish per-model prices on the release pages — check the ModelArk pricing page before you budget against any of them.
The GLM 5.3 Flash numbers are worth a second look
Zhipu claims GLM 5.3 Flash uses 3.01× less attention compute and 4.44× less KV cache than GLM-5.3 while keeping long-context performance. It also claims “native visual coding” — the model can inspect interfaces and rendered output inside the coding loop, then test and improve what it produced, across frontend work, games, Blender 3D and browser or GUI agent tasks. It is also stated to deliver finished PPTX, PDF, DOCX and XLSX files.
Those are specific, checkable-sounding numbers with no published methodology attached. Treat them as a vendor’s description of its own model until someone runs it independently.
dreamina-seedance-2-5-premium also remains whitelist-only. Source: the plugin’s own tagged README on GitHub.Seedance 2.5 Premium landed in the official ComfyUI plugin
On 29 September, BytePlus engineers tagged v0.2.0 of the official ComfyUI ModelArk plugin. This is where the video side of the month actually shipped.
What v0.2.0 adds
- Seedance 2.5 Premium (
dreamina-seedance-2-5-premium-260915): up to 30 seconds per generation. See the update note below — the accepted resolutions changed a day after this shipped. - Draft mode: render a 480p draft first, then generate the final video from that draft’s task ID — the obvious way to stop burning credits on 4K renders of a prompt that was not working.
- New Seedance parameters:
task_type(auto, reference, edit, extend) andoutput_format(mp4 or mov). - Seedream layer decomposition node: outputs a base image plus up to 16 alpha layers with accompanying layer JSON.
- Seedream 5.0 Pro updates: 1K / 1.5K / 2K presets, prompt optimisation set to standard or fast, PNG output with transparent background, and a new
reference_maskinput with amaskoutput. - Visual-understanding models added:
seed-1-8,seed-1-6andseed-1-6-flash.
The layer decomposition node is the one worth trying first. Getting a generated image back as sixteen separate alpha layers rather than one flat file is the difference between “nice picture” and something you can actually hand to a designer.
What did not happen
ModelArk’s Product updates page — the one that tracks platform features rather than models — still ends at August 2026. So September brought new models and new tooling, but no published platform changes. If you were waiting on a ModelArk feature, it has not shipped yet.
Separately, and worth knowing if you use Seedance through a third party: ElevenLabs has deprecated Seedance 1.5 Pro in its own Image & Video product, with generations stopping on 11 November 2026 and users pushed to Seedance 2.0. That is an ElevenLabs decision about which models it resells, not a BytePlus retirement, but the practical effect is the same if that is how you were reaching the model.
Should you care
Worth your time if
- You generate images at volume and have been paying Seedream 5.0 Pro rates for work that does not need Pro latency — Flash is the same feature set.
- You use ComfyUI and have been waiting for Seedance 2.5 Premium in an official node rather than a community wrapper.
- You want layered image output, which very few image APIs offer at all.
Skip it if
- You need published benchmarks before adopting a model. None of September’s additions have any.
- You were hoping for ModelArk platform features — nothing new since August.
If you want to try any of it, everything above runs on BytePlus ModelArk; the ComfyUI plugin is free and open on GitHub under the byteplus-sa organisation.
What we could not verify
Open questions
- Per-model pricing for Seedream 5.0 Flash. A price badge of $0.03 per image appears in a ComfyUI pull request, but we could not confirm it against BytePlus’ own pricing page, so we are not publishing it as fact.
- Any independent benchmark for DeepSeek V4.1 Flash, GLM 5.3 Flash or Seedance 2.5 Premium.
- Exact release dates. ModelArk dates its model releases by month (202609), not by day — so “September 2026” is as precise as anyone can honestly be.