v3.2.9 - Cerebras: gpt-oss-120b primary (fixes 404 on llama-3.3-70b + preps for Aug 17 glm-4.7 deprecation)

What's new in v3.2.9

Cerebras model migration ahead of their 2026-08-17 deprecation.

Cerebras (chat/refine primary)

• Retired llama-3.3-70b (has been 404-ing for weeks). Replaced with:
  • Primary: gpt-oss-120b — same 120B model as our Groq fallback, on Cerebras' faster infrastructure. Verified 200 OK end-to-end from the built dist.
  • Fallback: gemma-4-31b (Gemma 4)
• Retired zai-glm-4.7 (deprecating Aug 17). Config migration silently upgrades any existing user config naming these models to gpt-oss-120b on next launch.

Everything else unchanged

• Groq continues to serve: whisper-large-v3-turbo (STT), qwen/qwen3.6-27b (vision/OCR), openai/gpt-oss-120b (chat fallback)
• All features remain free-tier (Cerebras Developer + Groq Developer)
• 4 bundled API keys (2 Groq + 2 Cerebras) rotate on 429 rate limits

End-to-end verification of this dist

• verify_dist.py: 14/14 critical assets present + valid
• _bundled_keys.py shipped inside _internal/ (341 bytes, contains real gsk_ and csk- keys)
• Fresh launch: no "BUNDLED KEYS MISSING" warning
• Fresh launch: no Cerebras 404 (was constant before this release)
• Fresh launch: live POST api.cerebras.ai/v1/chat/completions HTTP/1.1 200 OK

Ship as Hotkeys-Windows.zip — extract anywhere, run Hotkeys.exe.