MAX_THINKING_TOKENS >= 64000 falls back to 21332 instead of using requested value

Resolved 💬 3 comments Opened Nov 7, 2025 by bukzor Closed Jan 9, 2026

Bug Description

When setting MAX_THINKING_TOKENS to 64000 or any higher value, the actual <max_thinking_length> value received in the system prompt falls back to 21332 instead of using the requested value.

Reproduction Steps

  1. Set MAX_THINKING_TOKENS=63999 - works correctly, value appears as 63999
  2. Set MAX_THINKING_TOKENS=64000 - fails, value falls back to 21332
  3. Set any value > 64000 - also falls back to 21332

Expected Behavior

Values >= 64000 should either:

  • Be accepted and passed through to <max_thinking_length>, OR
  • Show a clear error message indicating the maximum allowed value

Actual Behavior

Values >= 64000 silently fall back to 21332 with no error or warning.

Observed Values

Working values:

  • 62251 ✓
  • 63780 ✓
  • 63971 ✓
  • 63994 ✓
  • 63999 ✓

Failing values:

  • 64000 → falls back to 21332
  • Any value > 64000 → falls back to 21332

Notes

There appears to be a threshold check at exactly 64000 that's causing this fallback behavior. The value 21332 appears to be a hardcoded default.

View original on GitHub ↗

This issue has 3 comments on GitHub. Read the full discussion on GitHub ↗