DeepSeek V4 Flash responses become corrupted with repeated Q characters during long tool-assisted generations

Description

While using DeepSeek V4 Flash (Free) through OpenCode, model responses occasionally become corrupted during longer generations.
Instead of producing normal text, portions of the response are replaced with repeated Q characters and malformed words, making the output unreadable.
This appears to happen during or after reasoning/tool execution rather than being a rendering issue in the UI.

Environment

OpenCode: (fill in version/commit)
Model: DeepSeek V4 Flash (Free)
OS: Windows 11
Provider: (OpenRouter/OpenCode provider if applicable)

Steps to Reproduce

1. Start a conversation with DeepSeek V4 Flash (Free).
2. Give a relatively complex task that requires reasoning and code modifications.
3. Wait for the model to generate a long response.
4. Occasionally, the generated text becomes corrupted.

Expected Behavior

The model should produce coherent natural language throughout the response.
Example:

Placeholder: use the new 128×128 PNG for both themes...

Actual Behavior

The response becomes filled with repeated Q characters and corrupted words.
Example:

Q-placeholder QSS — low-Q one: embedded Q24×24 base64 PNG QSS app.js:44-45.
Not downloaded → QSS QSSQSSQSS no Q-Q asset.
...
Q migration QSSQ: check files exist...
...
QSS QSS values 44-45...

Later in the same response even PowerShell commands become corrupted:

Test-QPath
QReadQAllQBytes
Qictures
QHost

which makes the generated code unusable.

Frequency

encountered it today for the first time August 4 2026

Models used: deepseek-v4-flash-free

Plugins

No response

OpenCode version

1.18.11

Steps to reproduce

No response

Screenshot and/or share link

Operating System

Windows 11 25H2

Terminal

Windows Terminal (powershel 7)