跪拜 Guibai
← All articles
AI Coding · Doubao MarsCode · Artificial Intelligence

ByteDance's Trae Work Goes Free for All, Bundling Seed 2.1 Pro and a Full Coding Agent

By 甲维斯 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

A major Chinese tech company is giving away a capable coding agent with no apparent rate limits, which resets expectations for what free-tier AI coding tools should deliver. The bundling of document generation, data visualization, and autonomous coding into a single desktop app also signals that the agent market is consolidating around multi-modal workbenches rather than standalone code editors.

Summary

ByteDance's Trae Work desktop agent is now generally available with free access to the latest Seed 2.1 Pro model. The application splits into two modes: a Work mode for knowledge-intensive tasks like research reports, data mining, and paper summaries, and a Code mode that functions as an autonomous coding agent. Both modes are backed by generous token limits that allowed simultaneous multi-project testing without hitting rate limits.

The Work mode produced a 21-page PPT with ECharts visualizations from a short-video platform research prompt, generated a web-based AI market report with interactive tables, and created an illustrated paper review. The Code mode built a full multilingual learning platform over 68 minutes using React 18, TypeScript, Vite, and TailwindCSS, then automatically generated both a technical architecture document and a PRD.

Trae Work also includes mobile remote access with seamless conversation sync, and bundles access to multiple domestic models including GLM-5.2, Kimi-K2.7, and MiniMax M3. The international version of Trae still provides access to GPT and Gemini models without requiring a VPN, though Claude support was recently removed.

Takeaways
Trae Work provides free access to ByteDance's Seed 2.1 Pro model with no observed rate limits, even when running multiple coding projects simultaneously at peak hours.
The Work mode targets knowledge workers with tasks like research report generation, data mining with interactive charts, and paper summarization with auto-generated illustrations.
The Code mode built a 13-page React 18 + TypeScript + Vite + TailwindCSS multilingual learning platform in 68 minutes, including Zustand state management and mock data.
Generated projects include automatic quality checks: TypeScript type checking, production build verification, and responsive layout testing.
Each completed project produces two documents: a technical architecture document and a PRD requirements document.
Mobile remote access syncs conversations, task lists, and detailed dialogue information in real time without requiring a VPN.
The international Trae version still provides access to GPT and Gemini models from within China without a proxy, though Claude support has been removed.
Domestic model support includes GLM-5.2, Kimi-K2.7, and MiniMax M3 alongside ByteDance's own Seed series.
Conclusions

The free tier's token generosity is likely unsustainable and probably a pre-monetization growth tactic, given that ByteDance already charges for the international Trae version.

Trae Work's Work mode is positioned as a knowledge-work tool rather than a general-purpose chatbot, which differentiates it from consumer AI products like Doubao that handle everyday tasks.

The automatic generation of both technical architecture docs and PRDs after coding suggests ByteDance is targeting enterprise delivery workflows, not just prototyping.

Trae's ability to use GPT and Gemini models domestically without a VPN is an unusual regulatory workaround that gives it a distribution advantage over foreign competitors inside China.

Concepts & terms
Seed 2.1 Pro
ByteDance's latest large language model, positioned as a competitor to models like Claude Opus 4.7 on the ARENA benchmark.
Trae Work
ByteDance's desktop AI agent application that combines an office productivity mode (Work) with an autonomous coding mode (Code), plus mobile remote access.
Zustand
A small, fast state management library for React, used in the Trae Code-generated project to manage application state.
From the discussion

The disk footprint drew immediate attention, with one report of 8 GB for the agent and 3 GB for the editor, though another user saw a smaller install. Real-world testing over a month found Trae capable but requiring multiple rounds of iteration, while Codex delivered faster first-pass results. The free pricing is the main draw, but heavy queue times are eroding that advantage. For existing codebases, the tool struggles with holistic planning and burns through tokens, forcing a piecemeal interface-by-interface approach.

Install size reaches 8 GB for the agent and 3 GB for the editor, though the footprint may vary across environments.
Trae completes tasks from scratch but demands 2-3 rounds of correction, whereas Codex handles 90% of requirements in a single pass.
Free access is the primary reason for adoption, but growing queue times are making the free tier impractical.
On existing projects, the agent fails to produce a satisfactory plan even with detailed prompts, exhausting tokens and requiring manual interface-by-interface prompting.
Despite planning failures on legacy code, the tool still saves significant time when used as an assistant rather than a full-build driver.
Featured comments
心中要有光 1 likes

Tested it for a month, including a project built from scratch. Prompts were basically the same. Trae can get it done but needs 2-3 rounds of back-and-forth. Codex honestly nails 90% of requirements in one shot. But you can't beat Trae being free — it really saves money. The queue is getting ridiculous now though; the free tier is becoming unusable.

FlutterGo

If you start a project from zero and go step by step it's okay, but for an existing project, trying to get it to write a plan that meets expectations... I almost burned through all my tokens and still didn't finish. Now I have to find the code interfaces and have it fill them in one by one. It works as an assistant, but going straight for a full build — even when I describe the plan in great detail, it still makes mistakes. Still, it saves a lot of time.

See top comments, translated →
Source: juejin.cn ↗ Google Translate ↗ Backup ↗