At its annual Force conference, Volcengine released its newest base-model lineup, the Seed-2.1 series (Pro and Turbo). Starting from routine tests, work scenarios, and personal hobbies, we put Seed-2.1-Pro through a round of hands-on testing. One-line verdict: agents and coding have crossed the production-usability line, and multimodal recognition holds a surprise. Below are 6 test scenarios and prompts you can reuse directly.

Test 1: A Car Physics Animation (Overall Capability)

This test examines a model's physics modeling, seamless loop animation, spatial layering, aesthetics, and coding in one shot, with the constraint of vanilla JS only and no third-party libraries:

Implement a full-screen canvas in a single HTML file, without using any library. Simulate a moving car in side view as the main subject.
Keep the car visible in the foreground while the background scenery scrolls continuously, creating the sense that the car is driving forward.
Implement depth with a layered scene: the near ground, roadside elements, trees, utility poles, and distant hills or mountains
should move at different speeds to form a natural parallax effect. Animate the wheel rotation realistically, and add subtle body motion.
Let the environment flow smoothly past behind the car, with the scene repeating but staying varied. Use cinematic lighting and a unified sky to enhance the atmosphere.
The overall motion should feel calm, immersive, and real, with a seamless looping animation.

Seed 2.1 Pro's result slightly exceeded expectations: the background trees weirdly change height and the wheels sit a bit too high, but overall completeness is high, with no white screen.

Test 2: Front-End Marketing Site

Tests adherence to a visual style and component completeness:

Design and implement a homepage for an "AI financial forecasting platform" aimed at CFOs, including a hero section, trusted-customer logos,
3 core selling points, a live financial forecasting dashboard preview, customer testimonials, a pricing entry point, and a CTA.
Output a runnable single-file HTML/CSS/JS or a React component.
The visual style should feel premium in the vein of Linear/Stripe/Notion, without copying any brand outright.

The interface is clean and clear, prompt adherence is good, chart rendering looks great — solid and unremarkable overall, exactly as expected.

Test 3: Document-Driven Development (Spec Development)

Practitioners who are good at AI coding these days generally use document-driven development — have the AI generate a requirements document first, review it, then hand it back for development. The canonical case is building a CMS:

Read this PRD and build a CMS: [document URL]

Seed 2.1 Pro chose Next.js as the framework and SQLite as the database, and designed REST API endpoints as well. Development took about 5 minutes, create-read-update-delete all worked, and the features fully matched the document — beyond an unflashy interface, nothing to fault. Routine programming tasks can now be handed to it with confidence.

Test 4: Generating Headlines (with a Skill)

Coming up with titles for everyday articles is the most common task of all. Install the Xinzhiyuan headline Skill:

npx skills add joeseesun/qiaomu-xinzhiyuan-title

Then have Seed 2.1 Pro read an article link and generate 10 candidate headlines:

Use the Xinzhiyuan headline Skill to generate 10 candidate headlines; read the content here: [WeChat article URL]

Having the agent read Feishu and WeChat official account pages, then referencing the Skill's Reference documentation, makes headline quality more consistent than a bare prompt. Even the straight-laced Doubao can turn clickbait artist in a second.

Test 5: Competitive Research Report (with Tool Calls)

Research the official sites, pricing, core features, integration ecosystems, privacy compliance statements, and user reviews of 3 "AI meeting notes tools."
Output a competitive matrix, plus an MVP feature priority, differentiated positioning, and a 90-day roadmap aimed at SMBs.

Add "write it into a Feishu document" and it calls lark-doc precisely to write. Instruction following is good and every task got done; when direct scraping was blocked, it even fell back to Playwright to read the page.

Test 6: Photo-Based Fish ID (Multimodal, the Biggest Surprise)

As an avid angler, the author is building a personal iOS app: you take a photo, and the AI reads the EXIF to log the fishing spot while identifying the fish species and count. A previous test of Gemini 3.1 Flash misidentified a white strip as a "loach." Sent Seed 2.1 Pro an image with EXIF attached:

Read the geolocation information from this photo and identify how many fish are in it and what species they are

The result was touching: it called the EXIF tool to read the location (the Wenyu River), correctly identified the species and count (including two barely visible in the murky water), and even shared the Latin name and common names of the white strip.

Species: almost all white strip (Hemiculter leucisculus, colloquially can tiao, qing lin zi, you diao zi)

At least in this test scenario, multimodal recognition clearly outperformed Gemini 3.1 Flash.

How to Try It

Doubao Pro in Office mode, TRAE, TRAE WORK, and Coze all have Seed-2.1-Pro live. Enterprise and professional users generally connect via API and use it inside tools like Claude Code. To avoid interference from product system prompts and measure raw capability, all tests above were run with CC Switch + the Volcano Ark API, using cmux in the terminal. The Volcano Ark API is now fully open; apply directly and you can call it from Claude Code.

Every Skill used here is open source — just npx skills add and go. Seed 2.1 Pro's coding and agent-calling capability has reached the usability line: calling tools, writing documents, taking screenshots, and generating e-books all work, at a good price — for anyone dependent on AI workflows, the difference is very tangible.