---
source: 'https://howaiworks.ai/models/doubao-vision'
section: models
title: Seed1.5-VL
description: >-
  ByteDance Seed's standalone vision-language model: a 532M vision encoder with
  a 20B active-parameter MoE, top results on 38 of 60 public VLM benchmarks.
tags:
  - Seed
  - ByteDance
  - Vision Language Model
  - Multimodal Model
  - VLM
  - GUI Agent
  - Mixture of Experts
  - Chinese AI
category: Multimodal Models
developer: ByteDance
developerWebsite: 'https://www.bytedance.com/'
modelType: Vision-Language Model
releaseDate: '2025-05-13'
lastUpdated: '2026-07-08'
license: Proprietary
officialWebsite: 'https://seed.bytedance.com/en/public_papers/seed1-5-vl-technical-report'
docsPage: 'https://github.com/ByteDance-Seed/Seed1.5-VL'
pricingPage: 'https://www.volcengine.com/product/doubao'
---

# Seed1.5-VL

> ByteDance Seed's standalone vision-language model: a 532M vision encoder with a 20B active-parameter MoE, top results on 38 of 60 public VLM benchmarks.

## Overview

**Seed1.5-VL** is ByteDance Seed's standalone vision-language foundation model, described in its own technical report as a model *"designed to advance general-purpose multimodal understanding and reasoning."* It pairs a **532M-parameter vision encoder** with a **Mixture-of-Experts language model of 20B active parameters** — a deliberately compact design that nonetheless reaches state-of-the-art performance on **38 out of 60 public benchmarks**.

It is also, importantly, the *only* dedicated vision-language SKU in ByteDance Seed's model directory.

> **Status as of 2026-07-08: retired from Volcano Engine.** The model ID `doubao-1-5-thinking-vision-pro-250428` reached End of Service on **2026-03-31 at 14:00 (UTC+8)** and is no longer callable. Endpoints still pointing at it are automatically switched to `doubao-seed-2-0-lite-260215`. The technical report and cookbook remain valuable as reference material; the served model does not exist any more. See [Pricing & Access](#pricing-access) below.

### A note on "Seed 2.0 Vision"

This page previously described a model called **Doubao Seed 2.0 Vision**. **That model does not exist and never did.**

ByteDance's Seed 2.0 launch post enumerates the family exhaustively: three general-purpose agent models — Pro, Lite, and Mini — *"along with a dedicated Code model."* No Vision variant. The Seed 2.1 release that followed on June 23, 2026 shipped Seed 2.1 and Seed 2.1 Pro, again with no Vision variant. No `doubao-seed-2.0-vision` model ID exists on Volcano Engine or anywhere else.

The reason is straightforward: **vision is native to every Seed 2.x model.** [Seed 2.0 and Seed 2.1](https://howaiworks.ai/models/doubao-pro) take image and video input directly, and Seed 2.0 Lite was upgraded at the end of April 2026 to add audio, becoming the Seed series' first omni-modal understanding model. ByteDance had no reason to ship a separate vision SKU, and did not.

What ByteDance *does* list as a standalone vision-language model is Seed1.5-VL, and that is what this page now documents.

## Capabilities

From the technical report and the accompanying cookbook, Seed1.5-VL demonstrates strength in:

- **Complex visual reasoning** — including visual puzzles such as Rebus, which require inferring meaning rather than reading content off the image.
- **OCR** — text recognition and extraction from images.
- **Diagram understanding** — parsing charts, schematics, and structured graphics.
- **Visual grounding** — locating the specific region of an image that a phrase refers to.
- **3D spatial understanding** — reasoning about spatial relationships within a scene.
- **Video comprehension** — understanding temporal content, not just individual frames.
- **GUI control** — operating graphical interfaces as an agent, which the report singles out as an area of particular strength.
- **Gameplay** — agentic play, cited alongside GUI control as an agent-centric capability.

The technical report's own framing of the agent work is direct: *"in agent-centric tasks such as GUI control and gameplay, Seed1.5-VL outperforms leading multimodal systems, including OpenAI CUA and Claude 3.7."*

## Technical Specifications

- **Vision encoder**: 532M parameters
- **Language model**: [Mixture-of-Experts](https://howaiworks.ai/glossary/mixture-of-experts) with **20B active parameters**. Total parameter count not published.
- **Modalities**: image, video, and text input; text output
- **Volcano Engine model ID**: `doubao-1-5-thinking-vision-pro-250428` — **retired.** End of Service 2026-03-31 14:00 (UTC+8). No longer callable.
- **Technical report**: [arXiv:2505.07062](https://arxiv.org/abs/2505.07062), published May 2025
- **Weights**: closed. The companion GitHub repository is a cookbook under Apache-2.0 and does not host weights.
- **Context window**: **Not published by ByteDance.**
- **Knowledge cutoff**: **Not published by ByteDance.**
- **Pricing**: **Not published by ByteDance in English.**

The model ID's `250428` suffix indicates a snapshot dated 2025-04-28; the technical report followed in May 2025.

ByteDance describes the architecture as *"relatively compact"* and treats that compactness as the interesting result: the report is, by its own description, *"a comprehensive review of our experiences in building Seed1.5-VL across model design, data construction, and training at various stages."* It is written as a methods paper as much as a model card.

## Use Cases

- **GUI agents**: driving desktop and web interfaces from screenshots — the capability the report benchmarks against OpenAI CUA and Claude 3.7.
- **Document intelligence**: OCR plus diagram understanding on scanned documents, forms, and reports.
- **Visual grounding for downstream tools**: producing region references that a robotics or automation stack can act on.
- **Video understanding**: summarising, indexing, or answering questions about temporal content.
- **Chart and schematic analysis**: extracting quantitative meaning from data visualisations and technical drawings.
- **Multimodal reasoning research**: the compact architecture and the detailed training write-up make it a reference point for teams building their own VLMs.
- **Visual puzzle and reasoning evaluation**: Rebus-style tasks where the answer is not present in the image as text.

## Performance / Benchmarks

**State of the record**: ByteDance reports leadership claims and an aggregate count. It does not publish per-benchmark scores in its summary material — those live in the technical report itself.

- **38 out of 60 public benchmarks** — [state-of-the-art](https://howaiworks.ai/glossary/state-of-the-art-model) performance, per the technical report abstract. This is the model's headline claim.
- **GUI control and gameplay** — *"outperforms leading multimodal systems, including OpenAI CUA and Claude 3.7."*
- **Public VLM benchmark coverage** — the report states the model *"delivers strong performance across a wide spectrum of public VLM benchmarks and internal evaluation suites."*

We do not list individual benchmark scores here, because ByteDance's public-facing material does not carry them and we will not invent them. Readers who need the per-task numbers should read the [technical report](https://arxiv.org/abs/2505.07062) directly, where the full evaluation tables are presented.

Note that the OpenAI CUA and Claude 3.7 comparison dates from May 2025. Both comparison points have been superseded several times over; the claim describes the state of the field at publication, not today.

## Limitations

- **It is no longer served.** The Volcano Engine endpoint was taken out of service on 2026-03-31. This is the limitation that supersedes every other one below: you cannot call this model.
- **It is an older model.** Published in May 2025, Seed1.5-VL predates the entire Seed 2.x generation. For general multimodal work, [Seed 2.1](https://howaiworks.ai/models/doubao-pro) is the current flagship and handles vision natively.
- **No published per-benchmark scores in summary material.** The "38 of 60" figure is an aggregate; the individual results require reading the paper.
- **Stale comparison baselines.** The headline agentic comparison is against OpenAI CUA and Claude 3.7, both long since superseded.
- **No open weights.** The Apache-2.0 GitHub repository is a cookbook, not a weight release. Self-hosting is not possible.
- **No published context window or pricing.** ByteDance never published either, and the per-model Volcano Engine documentation page that once carried the endpoint limits now redirects to the general model list.
- **Volcano Engine dependency.** While it was served, access required a Volcano Engine API key, oriented toward the Chinese market. That dependency is now moot: the endpoint is gone and the weights were never released, so there is no way to run the model at all.
- **Compact by design.** A 20B active-parameter MoE trades peak capability for efficiency. On the hardest multimodal reasoning, larger frontier models will generally lead.
- **Content policy constraints.** As a Chinese-market model, outputs are shaped by content rules that differ from those governing US-hosted models.

## Pricing & Access

**The Volcano Engine endpoint for Seed1.5-VL has been retired.** The model ID `doubao-1-5-thinking-vision-pro-250428` is no longer callable, and it no longer appears anywhere in Volcano Engine's [模型列表](https://www.volcengine.com/docs/82379/1330310) ("model list").

### The retirement, as ByteDance documents it

Volcano Engine publishes a [模型下线公告](https://www.volcengine.com/docs/82379/1350667) ("model offline announcement"). Seed1.5-VL appears in 第六批模型下线说明 ("batch six offline notice"), whose schedule reads:

| Stage | Time (UTC+8) |
| --- | --- |
| 启动&通知 — start & notification | 2025-12-19 10:00:00 |
| EOM 模型停止新购 — end of marketing, no new provisioning | 2025-12-26 10:00:00 |
| EOS 模型服务下线 — end of service, model taken offline | **2026-03-31 14:00** |

The batch-six table lists the model with its replacement:

> 原文 (original): *"doubao-1-5-thinking-vision-pro-250428 | doubao-seed-2-0-lite-260215 | doubao-seed-2-0-lite-260215"*
>
> Translation: model ID | recommended migration target | system replacement if not migrated by the deadline — both columns naming `doubao-seed-2-0-lite-260215`.

The announcement defines the EOS milestone explicitly:

> 原文: *"EOS：End of Service & Support，模型停止服务，模型服务下线。"*
>
> Translation: *"EOS: End of Service & Support — the model stops serving; the model service is taken offline."*

And describes what happens to endpoints left pointing at a retired model:

> 原文: *"模型正式下线并残留接入点自动替换为新模型。"*
>
> Translation: *"The model is formally taken offline, and residual endpoints are automatically replaced with the new model."*

ByteDance frames this as routine rather than exceptional:

> 原文: *"一般情况下，每个版本的模型生命周期为3~6个月。"*
>
> Translation: *"In general, the lifecycle of each model version is 3 to 6 months."*

Seed1.5-VL was retired alongside its whole generation — `doubao-1-5-vision-pro-250328`, `doubao-1-5-vision-lite-250315`, `doubao-1-5-thinking-pro-250415`, `doubao-1-5-thinking-pro-m-250428`, and `doubao-1-5-ui-tars-250428` are in the same batch. No `doubao-1-5-thinking-*` model survives in the current list.

### Vision models that *do* appear in the current model list

From the 视觉理解能力 ("vision understanding") section of the model list, retrieved 2026-07-08:

**推荐模型 ("recommended models")**

- `doubao-seed-2-1-pro-260628`
- `doubao-seed-2-1-turbo-260628`
- `doubao-seed-evolving` (marked 快速迭代模型, "rapidly iterating model")

**往期模型 ("previous models")** — still listed, still callable

- `doubao-seed-2-0-pro-260215`
- `doubao-seed-2-0-lite-260428`, `doubao-seed-2-0-lite-260215`
- `doubao-seed-2-0-mini-260428`, `doubao-seed-2-0-mini-260215`
- `doubao-seed-2-0-code-preview-260215`
- `doubao-seed-character-260628`

Seven further entries carry the badge 即将下线 ("to be taken offline soon"), including the two older dedicated vision SKUs `doubao-seed-1-6-vision-250815` and `doubao-1-5-vision-pro-32k-250115`. Both are named in 第九批 ("batch nine") with an EOS of 2026-09-21 14:00 (UTC+8) and a migration target of `doubao-seed-2-0-lite-260428`.

Notably, `doubao-seed-1-6-vision-250815` is now the *only* model listed under GUI 任务处理能力 ("GUI task handling") — and it, too, is scheduled for retirement.

### What remains reachable

- **Cookbook repository** — [ByteDance-Seed/Seed1.5-VL](https://github.com/ByteDance-Seed/Seed1.5-VL) on GitHub, Apache-2.0, code samples and usage guidance, not weights. Its README still tells readers the model is deployed on Volcano Engine and invites them to "try it now"; that instruction is stale. The repository has had no push since June 2025.
- **Hugging Face Space** — [ByteDance-Seed/Seed1.5-VL](https://huggingface.co/spaces/ByteDance-Seed/Seed1.5-VL) still exists, but the Hugging Face API reported it in a `BUILD_ERROR` state when checked on 2026-07-08. Treat it as unavailable rather than as a working demo.

**No English-language rate card was ever published for this model**, and none exists now that it is retired. For rates on current models, consult the [模型价格](https://www.volcengine.com/docs/82379/1544106) ("model pricing") page.

## Ecosystem & Tools

- **[Volcano Engine (Ark)](https://www.volcengine.com/product/doubao)** — the API platform that hosted Seed1.5-VL until its retirement on 2026-03-31, and that hosts every current Seed model.
- **[Seed1.5-VL cookbook](https://github.com/ByteDance-Seed/Seed1.5-VL)** — official code samples and usage guide, Apache-2.0. Written against the now-retired endpoint.
- **[Seed 2.1 (Doubao)](https://howaiworks.ai/models/doubao-pro)** — the current flagship, with vision built in. The migration path in practice.
- **`doubao-seed-2-0-lite-260215`** — ByteDance's designated replacement for Seed1.5-VL, per the batch-six offline notice.
- **[Doubao Seed 2.0 Code](https://howaiworks.ai/models/doubao-seed-code)** — the dedicated coding model, which also carries multimodal perception.
- **[Seedream 5.0](https://howaiworks.ai/models/seedream)** — ByteDance's image *generation* line, as distinct from image understanding.

## Community & Resources

- [Seed1.5-VL Technical Report](https://seed.bytedance.com/en/public_papers/seed1-5-vl-technical-report) — ByteDance Seed's publication page
- [arXiv:2505.07062](https://arxiv.org/abs/2505.07062) — the full technical report, with complete evaluation tables
- [ByteDance-Seed/Seed1.5-VL](https://github.com/ByteDance-Seed/Seed1.5-VL) — official cookbook repository
- [ByteDance Seed Models directory](https://seed.bytedance.com/en/models) — canonical list of every Seed model, confirming Seed1.5-VL as the only standalone VL SKU
- [Seed 2.0 Official Launch](https://seed.bytedance.com/en/blog/seed-2-0-official-launch) — the post enumerating Pro / Lite / Mini / Code, with no Vision variant
- [模型下线公告](https://www.volcengine.com/docs/82379/1350667) — Volcano Engine's model offline announcements; Seed1.5-VL is in batch six, EOS 2026-03-31
- [模型列表](https://www.volcengine.com/docs/82379/1330310) — Volcano Engine's current model list, which no longer contains `doubao-1-5-thinking-vision-pro-250428`
- [Volcano Engine Doubao product page](https://www.volcengine.com/product/doubao) — API access and pricing

## Frequently Asked Questions

### Was there ever a "Doubao Seed 2.0 Vision" model?

No. ByteDance never released a model called Seed 2.0 Vision, and no `doubao-seed-2.0-vision` model ID exists. The Seed 2.0 launch post enumerates the family exhaustively as three general-purpose agent models — Pro, Lite, and Mini — "along with a dedicated Code model." There is no Vision variant, and none in Seed 2.1 either. Vision is native to every standard Seed model. If you encountered a page describing a "Seed 2.0 Vision," it was describing a model that does not exist.

### So where does vision live in the Seed family?

Inside the standard models. Seed 2.0 and Seed 2.1 accept image and video input directly, and Seed 2.0 Lite was upgraded at the end of April 2026 to add audio, becoming the family's first omni-modal understanding model. Vision is a property of the base models, not a separate SKU. Seed1.5-VL is the one standalone vision-language model ByteDance Seed lists in its model directory.

### What is Seed1.5-VL?

A vision-language foundation model from ByteDance Seed, described in its technical report as "designed to advance general-purpose multimodal understanding and reasoning." It pairs a 532M-parameter vision encoder with a Mixture-of-Experts language model of 20B active parameters.

### How large is Seed1.5-VL?

532M parameters in the vision encoder and 20B active parameters in the MoE language model. ByteDance calls this "relatively compact" and treats the compactness as a design goal rather than a compromise. The total (non-active) parameter count of the MoE has not been published.

### How does Seed1.5-VL perform on benchmarks?

ByteDance reports state-of-the-art performance on 38 out of 60 public benchmarks. In agent-centric tasks such as GUI control and gameplay, the technical report states that Seed1.5-VL "outperforms leading multimodal systems, including OpenAI CUA and Claude 3.7." ByteDance does not publish the per-benchmark scores in the summary material; they are in the technical report.

### What is Seed1.5-VL good at?

Complex visual reasoning including visual puzzles such as Rebus, OCR, diagram understanding, visual grounding, 3D spatial understanding, video comprehension, and GUI control.

### Is the model ID `doubao-1-5-thinking-vision-pro-250428` still callable?

No. Volcano Engine's 模型下线公告 ("model offline announcement") lists it under 第六批模型下线说明 ("batch six offline notice") with an End of Service date of 2026-03-31 14:00 (UTC+8). That date has passed, and the ID no longer appears anywhere in Volcano Engine's current 模型列表 ("model list"). Endpoints still pointing at it are automatically switched to `doubao-seed-2-0-lite-260215`.

### How do I access Seed1.5-VL?

As of 2026-03-31 you largely cannot. The Volcano Engine model ID `doubao-1-5-thinking-vision-pro-250428` reached End of Service and is no longer callable; ByteDance's designated replacement is `doubao-seed-2-0-lite-260215`. The Apache-2.0 cookbook repository is still online (its README still advertises the retired ID), and the Hugging Face Space still exists, though Hugging Face reports it in a build-error state as of 2026-07-08. For current vision work, use a listed model such as `doubao-seed-2-1-pro-260628`.

### Are the Seed1.5-VL weights open?

No. The model is proprietary and was served through Volcano Engine until its retirement on 2026-03-31. The companion GitHub repository is a cookbook and usage guide released under Apache-2.0 — it does not host model weights. Because the weights were never released and the endpoint is gone, the model is now effectively unavailable.

### When was Seed1.5-VL released, and when was it retired?

The technical report was published in May 2025 (arXiv:2505.07062). The Volcano Engine model snapshot carries the date 2025-04-28, visible in the model ID `doubao-1-5-thinking-vision-pro-250428`. Volcano Engine began the retirement on 2025-12-19, stopped new provisioning (EOM) on 2025-12-26, and took the model out of service (EOS) on 2026-03-31 14:00 (UTC+8).

### Should I use Seed1.5-VL or Seed 2.1 for vision work?

Seed 2.1, because Seed1.5-VL is no longer served. Seed 2.1 is the current flagship and handles image and video input natively. Seed1.5-VL remains interesting as a research reference — a dedicated, compact vision-language model with a detailed training write-up, particularly for GUI-agent and grounding workloads — but it can no longer be called on Volcano Engine.

## Related

### Related models

- [Seed 2.1 (Doubao)](https://howaiworks.ai/models/doubao-pro)
- [Doubao Seed 2.0 Code](https://howaiworks.ai/models/doubao-seed-code)
- [Seedream 5.0](https://howaiworks.ai/models/seedream)
- [Gemini 3.5](https://howaiworks.ai/models/gemini)
- [GPT-5.6](https://howaiworks.ai/models/gpt)
- [Qwen3.7-Max](https://howaiworks.ai/models/qwen)

---

Source: https://howaiworks.ai/models/doubao-vision — HowAIWorks.ai