Some notes from the 2025 Volcano Engine Force Conference

Overall impressions after watching

ByteDance is pushing hard on multiple fronts.

It pushes every form to the extreme at the product level. Not only in multimodal conversational assistants (to C), cloud native, FaaS and other areas — at the entry point and the ecosystem level it’s already cruising on the highway.

What struck me most was Trae Solo. Even though it’s a futures product, so I won’t pay much attention to it for now, still —

why does Trae, an app aimed at programmers, get so many sessions and so much airtime over two days? I think it’s wildly ambitious.

Trae Solo could one day become another form of browser, or rather another form of integrated office software.
I say it replaces the browser because it will become the entry point for acquiring all information.
I say it becomes office software because it will become the production tool for producing every kind of digital output.

For me, trae can replace Doubao, can replace Coze Space, can replace manus, can replace fellou, can replace every agent. Thanks to MCP it has unmatched extensibility; thanks to being a VSCode environment itself it naturally has the underlying ability to run all kinds of local tools; and thanks to code writing and compilation it can directly serve as the final fallback that none of the above can achieve.

It can produce code, produce documents, and analyze data. It can be a work partner and expert in any form.

I’ve also used Trae before to build math practice problems and English vocabulary memorization for kids — automatic question generation and grading, with scores sent to Feishu — all done in a few sentences. This thing is really good, it benefits everyone. Everyday needs no longer require knowing how to program.

Doubao Large Model 1.6 released

This time the name carries “Seed”, which I think signals the start of a new chapter for Doubao in the new AGI era — AGI emphasizes multimodality. Volcano’s solutions have always led the industry in multimodality. Last year Doubao’s voice interaction was already natural and smooth, and at the time it felt like it had caught up with ChatGPT, which was quite a shock to me.

The whole series supports multimodal input and 256K context, including three versions: doubao-seed-1.6, 1.6-thinking, and 1.6-flash. Among them, 1.6-thinking stands out in math and logical reasoning, reportedly reaching the level of scoring 700+ on the Gaokao.

DeepResearch feature: Can complete deep research tasks in 5–30 minutes, competing directly with products like Fellou and Manus. The feature combines multi-round thinking with real-time search to analyze and solve complex problems.

Improved multimodal capabilities:

• Video chat supports automatically recognizing on-screen content, meeting context-dependent user filtering needs
• Enhanced GUI operation capabilities, seemingly integrating UI Tars technology
• Supports input and processing of images, voice, music, and other modalities

Revolutionary pricing model: The first to price by “input length” tiers, with deep thinking, multimodal capabilities and the base language model all priced the same:

• 0-32K input tier: ¥0.8 per million tokens input, ¥8 per million tokens output
• 32-128K input tier: ¥1.2 per million tokens input, ¥16 per million tokens output
• 128-256K input tier: ¥2.4 per million tokens input, ¥24 per million tokens output

Calculated at a 3:1 input-to-output ratio, the overall cost is only one third that of the Doubao 1.5 deep-thinking model or DeepSeek R1, greatly lowering the barrier to enterprise applications.

$$ 3/4*0.8 + 8/4 = 2.6¥ $$

Trae

Bringing technology to everyone.

Solo
The Solo Web App Builder demonstrated supports voice input, offering a convenient way for multi-role collaboration among product managers, developers and others.

The multi-role pairing feature lets product managers quickly input requirements by voice, while developers see them in real time and turn them into technical implementations, significantly improving collaboration efficiency.

Futures — just wait.

Video generation

Volcano Engine’s Seedance 1.0 series of video generation models brings breakthrough content creation capabilities, including two versions: Seedance 1.0 Pro and Seedance 1.0 Lite.

Seedance 1.0 Pro technical features:

• Multimodal input support: Accepts text descriptions and image inputs, and can generate 1080P high-quality videos with seamless multi-shot switching
• Industry-leading performance: Ranks first in text-to-video and image-to-video tasks, especially good at building complex scenes and generating character actions
• Precise instruction following: Understands fine-grained shooting requirements, supports cinematic camera movement and sensible motion interaction design
• Open platform access: Services are already open to the public via the Doubao App, Jimeng AI, Volcano Engine and other platforms

Seedance 1.0 Lite version:

• A lightweight video generation solution supporting basic text-to-video and image-to-video functions
• Optimized for generation speed and resource usage, suitable for rapid prototyping and lightweight application scenarios

Disruptive pricing strategy:

• Seedance 1.0 Pro costs only ¥0.015 per thousand tokens; generating a 5-second 1080P video costs about ¥3.67
• At that rate, a budget of ¥10,000 can generate about 2,725 5-second 1080P videos, dramatically lowering the barrier to producing high-quality video content

The deep-thinking video generation capability combines the logical reasoning of the Doubao large model with Seedance’s visual generation, automatically analyzing text requirements, planning shot sequences, and generating video content that follows a narrative logic.

Doubao real-time voice model fully launched

Developers can now call it
Dialects
Soft voice
Singing
Podcast generation: from prompts, links, long-form text content

DeepSearch, Agent

It’s actually an application in Volcano Ark — not on Doubao, and not on Coze Space??
The features are exactly the same as Fellou, but you can manually add MCP, and the generated outputs can be more varied.
OK fine, for me I’ll still implement it with Trae + MCP.

When generating web pages you can check Canvas to view a live preview.

veFaaS MCP can be deployed on the public network; Volcano provides veFaaS MCP.

Making videos
Trae Viral Video Creator: 😐 this thing was created by myself, and it needs to be paired with vevod MCP

Detailed steps are here: https://www.volcengine.com/docs/82379/1592795

• Agent Development Guide: “Xianyanbao” market strategy analysis and report generation
• Agent Development Guide: generating the “Xianyanbao” website
• Agent Development Guide: creating the “Xianyanbao” marketing video

PromptPilot

Volcano Engine’s PromptPilot is an intelligent prompt optimization platform designed specifically for large AI models, dedicated to solving the core pain point of the high barrier to writing prompts. It offers a one-stop automated solution covering the full lifecycle from prompt generation, debugging, optimization to iteration.

Core features:

• Requirement translator: Precisely captures user intent through interaction, turning vague requirements into clear instructions
• Intelligent optimization: Uses a simplified “multiple-choice” mode, automatically generating diverse answer options to guide users toward clarifying their needs
• Complex task decomposition: Performs especially well for image and video scenarios, automatically breaking multimodal tasks into multiple steps and searching for the optimal path
• Closed-loop optimization system: Lets users flag Bad Cases, turning problem cases into data assets and driving continuous model evolution
• Reflection and iteration capability: Capable of simulating human thinking — reflection, summarization, and error analysis — and finding through multiple rounds of automatic iteration the “golden prompt” that reliably produces the best results

Developer support:

• case radar: Helps developers accumulate case assets
• Prompt SDK integration: You can integrate the SDK into your app to keep prompt data flowing back
• Supports web search and knowledge bases: Can optimize prompts by combining real-time web information and custom domain knowledge bases

The visual understanding solution demonstrated PromptPilot’s power: automatically design a scheme to count vehicles in a parking lot, generate a solution, and keep validating and optimizing it with new samples. Its ability to generate code shows it has moved beyond the scope of a traditional prompt tool to become a bridge between natural language and machine execution.

Implementation

Knowledge product management

Reinforcement learning
The veRL framework has been open-sourced.

Multimodal data lake
Vectorization, search by image, export to Volcano Ark

AI data expert, a data agent product.

• Knowledge bases, SQL, meeting minutes, etc. — results within an hour.
• An actionable marketing plan within 5 hours.

AgentKit: integrates all the infrastructure mentioned earlier
TrainingKit: 30% better than open-source solutions, integrates veRL
ServingKit: inference optimization.
Aren’t these already available directly with Volcano Ark? It’s just suited to customers who want to integrate atomic capabilities — building their own large models and deploying their own models.

Customer: Weibo Zhi Sou. Haha, the Weibo person told the plain truth: DeepSeek blew up the public discourse, but the deployment cost is unbearable.

veOmniverse actually partnered. Nice.

Afternoon

HiAgent has gotten impressive,
learning while doing, online learning (under the hood it should be using the prompt SDK)
automatic, fast planning and setup.

HiAgent agent case studies:
Nankai University: research agent — research intelligence
Xi’an Jiaotong University: oral imaging interpretation

the first year of agents

Security, but if the customer stubbornly insists on private deployment there’s nothing you can do

Breakout sessions

Multimodal data lake

Volcano Engine’s multimodal data lake is one of the core components of its AI cloud-native full-stack service, designed specifically for processing and managing large-scale multimodal data. Its core features include:

• Unified multimodal data storage: Supports integrated storage and management of text, images, audio, video, and other data types
• Intelligent processing: Built-in automatic vectorization converts unstructured data into high-dimensional vectors, enabling efficient similarity retrieval
• Cross-modal retrieval: Enables cross-modal queries such as “search by image” and “search by text”
• Works with large models: Seamlessly connects to AI capabilities such as the Doubao large model, and supports pulling data directly from the data lake for model training or inference
• Volcano Ark integration: Supports exporting processed data to the Volcano Ark platform to build enterprise-level knowledge bases

Typical application scenarios:

• UGC content analysis and management: unified management and intelligent retrieval of images, videos, and text on social media platforms
• Enterprise knowledge base construction: integrating multimodal knowledge assets such as documents, meeting recordings, and product images
• Intelligent media production: supports intelligent tagging, retrieval, and recombination of video assets
• Multimodal training data preparation: provides structured multimodal datasets for AI model training

LanceDB

Volcano Engine has integrated LanceDB (one of the most popular open-source vector database formats today), giving the multimodal data lake efficient vector storage and retrieval. LanceDB uses a columnar storage format and supports high-concurrency queries and dynamic data updates, making it especially well suited to large-scale AI application scenarios.

The underlying principles of the multimodal data lake involve vector embedding technology, distributed storage architecture, and cross-modal data association algorithms. By mapping different types of data into a unified vector space, it achieves efficient storage, retrieval, and analysis. This technology gives enterprises a powerful solution for handling ever-growing unstructured data.

Multimodal data governance

Video, annotation
Processing PB-scale data…
Nemo Curator
Video annotation model optimization

Pretty hardcore too, but I don’t need it for now

Data agent: an all-scenario revolution

The best time to embrace Agentic Data.
Hmm, I don’t think it’s mature yet. Mainly because I don’t have that much data to test with, so I can’t appreciate the value of its case studies at all. And I have no idea what kind of hallucinations and slip-ups it will have.

Ark

prompt tuning

trae

trae has a beta version, one step further into the future

Coze

agent low-code, full-code building
Evaluation platform

Eino

This one you can really evaluate yourself

This year there’s some conceptual design in everything:
Accumulating cases, growth loop
Prompt engineering tuning

Deerflow can be experienced for free on Ark
UI-Tars

Welcome to follow the Volcano Engine Force Conference Developer Forum! There are still 3 great breakout forums this afternoon — if you’re interested, check them out 👇👇👇

Volcano Ark Developer Day: https://www.volcengine.com/live/event/force-2506-developer-fangzhou
Coze Enterprise Exchange Day: https://www.volcengine.com/live/event/force-2506-developer-coze
ByteDance Open Source Developer Day: https://www.volcengine.com/live/event/force-2506-developer-opensource

Open source session

KV cache
PD disaggregation

https://deerflow.tech Deep research
Based on LongChain,

demo:
Ask a question directly
Add [Background Research]: first investigate the question, then decompose the task — put the search results into the user prompt, and you instantly get better context.
[Magic Wand]: prompt optimization.
[edit plan]: propose changes

!! Never rely on a large model to do any math calculations;
Strategy: use Python to generate a program, then run the program to compute the math.

report, further editing
Supports MCP

Deployment method: Volcano Engine FaaS demo, deploy directly

A LangChain promoter, haha — one line of code to implement memory, two lines to integrate MCP
https://www.langchain.com/

Layers:
Layer one, intent recognition, Coordinator — recognizes the intent of the question at extremely low cost; if it’s not necessary, it doesn’t enter the large-model answering flow.
Layer two, Planner — generates a plan based on the user’s semantics.
Research Team — mechanically executes exactly according to the Planner’s plan. This layer doesn’t need a large model, because the large model would just follow the plan anyway — a total waste.
researcher: includes calling browser use
coder: writes code
The Planner judges the results from the research team, checks whether the plan is satisfied, and aggregates the multiple planned tasks across all teams.