An article explaining how to use Talos: how to choose models, deliver files, operate the browser, implement four scenarios, and a three-step guide to get started.
Project Background
Talos is quick to get started with, but long-term users will notice that true proficiency comes down to a few key things: when to switch models, how to deliver files, and how to make the browser perform tasks for you. This knowledge is scattered across changelogs and case studies, and new users often spend significant time piecing it together.
So we created a product user manual, bringing together features, parameters, scenarios, and usage suggestions in one place, organized from capabilities to scenarios to techniques. Below is a distilled version from the manual—ready to use right after reading.
What is Talos
Talos is an out-of-the-box AI Agent. You give it a task in a single sentence, and it goes off to research, create documents, draw images, and write code—handling everything from work tasks to personal creative projects.
What sets it apart from a chat window is the deliverable: when a task ends, you receive a file you can open, forward, or continue editing, rather than a reply you need to reorganize on your own.
Getting Started in Three Steps
Step 1: Download and Log In
Download the installer from the official website, available for both Windows and macOS. Log in after installation. Major domestic and international models are ready to use out of the box, with no additional network configuration needed.
Step 2: Start a New Session
Each session corresponds to one task. Create one for competitive research and another for writing an article. Sessions are independent of each other, and both history and outputs are archived per session.
Step 3: Describe the Task Clearly
A single-sentence task will get things started, but specifying three things greatly improves results: what result you want, who the audience is, and any formatting requirements. For example, 'Break down this 57-page courseware into 11 knowledge points, provide a note and a card for each, and finally assemble an overview map' is far more effective than 'Help me organize the courseware.'
Four Core Capabilities
1. Models: Choose by Task, Not by Price
Talos currently integrates 10 text models, 5 image models, and 5 video models. Text models generally fall into three tiers:
| Tier | Representative Model | Best For |
|---|---|---|
| Most Capable | sonnet-5.5 | Writing documents, conducting analysis, modifying code |
| High Cost-Efficiency | deepseek-flash, gemini-3.8-flash, glm-5-3 | Batch data cleaning, format conversion, field extraction |
| Lightweight | gemini-3.5-flash-lite | Single-sentence Q&A, typo correction, tagging and classification |
Among image models, gpt-image-2.5-sunburst is the default: supports 1K or 2K resolution, five quality levels, transparent backgrounds, and up to 4 images at once. Stick with it if you're unsure. For large-format, image editing, or maintaining character consistency, switch to gemini-3.1-flash-image, which supports up to 4K.
The default video model is Seedance 2.0, with options for 5s or 10s, and resolutions from 480p to 1080p. Veo 3.1 supports 4, 6, or 8 seconds and aspect ratios of 16:9 or 9:16.
Another often-overlooked parameter: Thinking Effort. Set it low for simple tasks to save time and cost; medium for everyday coding; high for multi-file refactoring, stubborn bugs, or architectural decisions.
2. Files: Upload, Analyze, and Deliver
Files can be dropped directly into the session window—PDF, Word, PPT, images, and spreadsheets are all readable. Outputs are similarly delivered as files. Talos presents the file directly to you, with support for copying and sharing.
There is also a dedicated file delivery portal that aggregates all outputs from historical sessions, essentially giving you a continuously growing library of finished products.
3. Browser: Access the Real Internet
The built-in browser can open webpages, read content, perform clicks, navigate pages for data collection, and download files. For sites requiring login, it hands the page over for you to manually complete the step, then resumes execution.
A typical use case is competitive research: open a store page, capture product titles, prices, sales volumes, and review counts, compile them into a structured table, and then conduct analysis. You can see everything it's doing throughout the process.
4. Memory: Learns Your Preferences Over Time
Talos records three types of information: site structures, personal preferences, and reusable workflows, all saved locally. It remembers long-term preferences, project backgrounds, and hard constraints, but not temporary requirements specific to a single task.
This means the second time you perform a similar task, you don't need to explain the background all over again.
Four Typical Scenarios
Learning
Reading papers, converting courseware, organizing videos, writing lab reports. A single-sentence task produces reviewable material, not just an explanation.
E-commerce
Product selection research, competitive analysis, negative review mining, listing copywriting. A typical task involves opening platform pages, scraping data, making comparisons, and generating reports.
Self-Media / Content Creation
A streamlined pipeline from topic selection to finished product:
| Task | Input | Output |
|---|---|---|
| Product Promo Video | A one-sentence requirement and product screenshots | A 52-second 1080p finished video with subtitles and music |
| Graphic Marketing Materials | A usage scenario or product selling point | A complete set of graphic cards with accompanying copy |
| Long Video Clips | A single 17-minute video | 578 timestamped captions, an SRT file, and 2 clips with sound |
| Industry Content | A one-sentence requirement | An 18-page editable PPT, a PDF of the same version, and speaker notes |
Design and Creative Work
Game art assets, character and scene illustrations, 3D furniture and product visuals, shoe designs, and posters. The emphasis is on style consistency: generating assets in batches under the same description instead of tweaking images one by one.
Three Usage Suggestions
Lead with the Conclusion
When requesting a report or proposal, state the conclusion first, then provide evidence. You can directly write in your prompt: 'Give me a one-sentence judgment first, then list three supporting reasons,' and the readability of the output will be much higher.
Avoid Switching Models Mid-Task in Long Sessions
Switching models resets the prompt cache, requiring the entire conversation history to be re-processed at full cost in the next round, which can significantly increase expenses. If you need to switch, do it in a new session or very early in the conversation.
Use Examples to Ask Questions
If you're unsure how to phrase a request, directly paste in a similar case study or a sample output and ask it to follow that format. This is more effective than repeatedly describing the desired format.
FAQ
Q: Do I need to set up a separate network environment?
A: No. Log in after installation, and you can use it immediately. Major domestic and international models are already integrated.
Q: Are files uploaded to the cloud?
A: Site structures, personal preferences, and workflow records are saved locally. Content that requires model processing is sent along with the request to your chosen model service.
Q: Where can I find the output files?
A: You can open them directly in the current session's directory. All historical outputs can be viewed uniformly in the file delivery portal.
Q: What if more models make it harder to choose?
A: Select by task tier: use the most capable tier for quality, the lightweight tier for speed, and the high cost-efficiency tier for batch processing. When in doubt, run a round with the default model first and switch based on the results.