Repository navigation
GLM models excluded from reasoning variants — no way to toggle thinking at runtime #18598
Description
Description
GLM models (GLM-5, GLM-4.7, etc.) are explicitly excluded from reasoning variants in ProviderTransform.variants(), which means users cannot toggle thinking mode on/off during a session — neither through the UI variant selector nor via SDK variant field.
This is limiting because GLM-5's thinking mode is expensive for simple tasks and Z.AI recommends dynamically switching between thinking/non-thinking modes depending on task complexity.
Static config (model.options) works but requires restarting the session. Dynamic toggling via variants does not work because variants() returns {} for GLM.
Steps to reproduce
- Configure GLM-5 model with
"reasoning": true - Open a session
- Try to switch reasoning variant — no options available
Expected behavior
GLM models should have none/high reasoning variants, allowing users to toggle thinking on/off at runtime, similar to how grok has low/high.
- I have verified that this issue has not already been requested
- I have read the contributing guidelines
Reacted by Sophie, Gabriel Patzleiner, Konstantin Scherban, Evan Du and Tristan Druyen
Metadata
Metadata
Assignees
Labels
No labels
Activity
github-actions commentedon Mar 22, 2026
This issue might be a duplicate of existing issues. Please check:
If your issue is distinct from these, please clarify how it differs.
feat(core): add reasoning variant toggle for GLM models
0717376 commentedon Mar 22, 2026
This issue is related but distinct from both:
[FEATURE]: Add disable_reasoning parameter for GLM/Cerebras thinking toggle #11987 proposes a new
disable_reasoningconfig field at the agent level. Our issue is about the existingvariants()mechanism — GLM is excluded from it, so runtime toggling via the UI variant selector doesn't work. The fix (PR feat(core): add reasoning variant toggle for GLM models #18599) uses the existing variants infrastructure, no new config fields needed.[FEATURE]: Add thinking/reasoning support for GLM-4.7, Kimi K2.5, and MiniMax M2.1 on Zen #12396 is about Zen gateway rejecting thinking config for paid models. Our issue is about the core
ProviderTransform.variants()excluding GLM entirely, regardless of provider.rekram1-node commentedon Jun 20, 2026
this is intentional, aside from glm 5.2 which actually supports them.