Ask an everyday question or describe the result you want in your own words. I answer directly, find the right tool when needed, and prepare the action for your approval.
BYTEQUANT AI · RUNS ON YOUR DEVICE
Chat naturally, then put the right tools to work
Ask an everyday question, describe a task, or edit an image on your device. ByteQuant AI remembers context, selects suitable tools, and completes supported work in the same screen. A device-suited open-source model is prepared only with your approval when you want more natural conversation.
BQ
Hi, I’m ByteQuant AI.ByteQuant AI 7.7 · ByteQuant AI · On your device
ByteQuant AI · On your device
What should we solve together today?
Write in everyday language. I answer directly first; if action is needed, I show the right tool and prepared input.
Everyday questionsWriting and decisionsFiles, images, and 342 tools
ByteQuant AI · On your deviceOptional; fast answers work without a download+
Download-free instant AI is readyChat with a real open-source model without sending messages to a remote AI service.
✓ Matching and automation for 342 tools✓ Tab-scoped conversation context↓ Downloaded once on first use
Model and device options+
ⓘ First use downloads your chosen model once. Your conversation is not added to the download request or sent to a remote model. ~0.4 GB
Qwen3 · Apache-2.0 · Web Worker · no remote inference◉ This conversation stays in this tab · Context on
Find a tool or explain an error+
BYTEQUANT AI 7.7
Try 7.7: describe the task, send the data next
The fast response layer executes specific calculations and text operations using explicit rules. These examples work without downloading a model. Each includes an expected result so you can check the output instead of relying on a success message.
Table to JSON
- 1. Your request
- Convert CSV to JSON
- 2. Your next message
id;name 007;Ada
- Result to check
"id": "007"
The first message selects the operation. Paste the table in the second message to run that same plan. Semicolons and tabs copied from spreadsheets are supported. 007 remains a string, preserving its leading zeros. Repeated column names stop conversion: give each column a unique name and retry.
Open the related tool →Encode, then decode
- 1. Your request
- Base64 encode: Merhaba
- 2. Your next message
Now decode this
- Result to check
Merhaba
The first output is TWVyaGFiYQ==. The second request connects that output to the active Base64 operation. Compare the decoded text with the original. Base64 is not encryption and does not protect secrets. Starting a separate JWT task does not silently reuse the preceding output as a JWT.
Open the related tool →Change a calculation with a short reply
- 1. Your request
- 15% of 200
- 2. Your next message
What about 30%?
- Result to check
60
The first answer should be 30; the second should be 60. Only the percentage changes, while the base remains 200. The calculation is 200 × 30 ÷ 100. Support for this short percentage correction does not imply that arbitrary mathematical problems can be solved.
Open the related tool →Easier input: use named fields for CAGR, ROAS/ROI, Content-Disposition, PICO and date addition. CSV pivot and JSON sorting rules also have fields. Raw input remains available under advanced editing. Duplicate keys or unrecognised lines prevent guessed field mapping; your original text stays intact.
BYTEQUANT VISUAL
Edit an image or create a new visual draft
Upload an image and describe the edit, or create an original vector draft from a prompt. Work starts only in this tab.
LOCAL-FIRST
What the Agent does—and does not do
It does not replace the tools. It makes the right tool and a safer order easier to find.
Understands the need
Matches the outcome, file format, and privacy signals.
Builds a short plan
Orders suitable tools, explains why, and asks for missing details.
Keeps you in control
File selection, execution, and downloads happen only when you act.
Explains the next move
Turns common errors into plain language and offers nearby alternatives.
REAL USE CASES
How one sentence becomes a usable result
These examples show what the download-free fast layer and tool automation actually do. ByteQuant AI does not invent live facts; it works with the text, choices, or files you provide.
Writing and communication
“Write a short apology email for a delayed delivery.”
Produces a sendable draft with accountability, a revised date, and a corrective action; you can then make the tone warmer or more formal.
Decision support
“Compare freelance and full-time work by risk.”
Separates domain-specific criteria such as income stability, autonomy, benefits, client acquisition, and financial runway, then suggests a reversible test.
Error explanation
“What does TypeError: cannot read properties of undefined mean?”
Explains the cause in plain language, shows safe access, and asks only for the code context needed for a concrete fix.
Tool automation
“Combine these photos into one ordered PDF.”
Selects the image-to-PDF tool, explains the file requirement, prepares the handoff, and waits for approval before processing.
PRIVACY BY DESIGN
Your content stays on your device
bytequant.org · HTTPS · ByteQuant AI 7.7
- ✓
Planning, automation, and on-device inference do not send goals or tool inputs to a ByteQuant server.
- ✓
Short conversation memory stays in sessionStorage for the open tab only.
- ✓
Optional open-source model assets download only with your approval; chat content is not added to that request.
- ✓
AI answers can be wrong and are not legal, medical, security, or identity-verification decisions.
FAQ
Short, clear answers
Is ByteQuant AI real AI?+
Yes. Its explainable tool-planning and verification core works on every device; on supported devices, your selected Qwen3 model generates answers inside a Web Worker. Chat is not sent to a remote AI API.
Does it remember a conversation?+
It uses recent goals and answers as context in the same tab. Closing the tab ends this temporary memory.
Can it run tools automatically?+
It can complete supported bounded text workflows, show the result in the conversation, prepare input in the first tool, and send a plan to Workstation. File selection, passwords, code execution, downloads, and high-impact decisions require your approval.
Does it really edit images?+
Yes. It applies resizing, brightness, contrast, saturation, grayscale, blur, rotation, and flipping to PNG, JPG, and WebP with Canvas or OffscreenCanvas on your device. It can create prompt-based SVG drafts, but does not claim object removal or photorealistic generation.
Why does it sometimes ask before downloading a model?+
Core planning and automation start immediately. When a device-suited open-source model is needed for more natural replies, ByteQuant shows the size first and downloads only with your approval; a cached model is reused automatically.