# Changelog v0.5.20 - 2026-06-10

### Reverted
- **groq**: Reverted 7 non-coding models incorrectly added in v0.5.19:
  - `whisper-large-v3`, `whisper-large-v3-turbo` (audio transcription — not coding)
  - `canopylabs/orpheus-arabic-saudi`, `canopylabs/orpheus-v1-english` (speech — not coding)
  - `meta-llama/llama-prompt-guard-2-22m`, `meta-llama/llama-prompt-guard-2-86m` (security guardrails — not coding)
  - `openai/gpt-oss-safeguard-20b` (safety — not coding)
  - groq remains at 8 coding models; total catalog back to 154 models.

### Kept (from v0.5.19 audit — all valid coding changes)
- **nvidia**: Added `nvidia/nemotron-3-ultra-550b-a55b` (S+, 1M context, NVIDIA NIM)
- **nvidia**: Corrected context windows:
  - `deepseek-ai/deepseek-v4-pro`: 128k → 1M
  - `deepseek-ai/deepseek-v4-flash`: 128k → 1M
  - `mistralai/mistral-small-4-119b-2603`: 128k → 256k
- **nvidia**: Removed deprecated model `z-ai/glm5` (replaced by `z-ai/glm-5.1`)

### Notes
- This catalog focuses exclusively on **coding LLMs** — audio, speech, guardrail, and safety models are out of scope.
- Model counts: nvidia 27 (was 27 → no net change after removing glm5 and adding nemotron-ultra), groq back to 8.
- audit_state.json fingerprints corrected for groq.