v0.3.0

Added

• mixtape-server (experimental): HTTP server with AG-UI protocol support. API surface may change in future releases.
• Claude Opus 4.6 model (flagship, 200K context, 128K output)
• Claude Opus 4.1 model (200K context, 32K output)
• Nova 2 Sonic model (1M context, 65K output)
• Mistral Ministral 3B, 8B, 14B models
• Mistral Pixtral Large model
• Mistral Voxtral Mini 3B and Voxtral Small 24B models (speech+text input)
• Qwen3 32B, Qwen3 Coder 30B, Qwen3 Next 80B, Qwen3 VL 235B models
• Google Gemma 3 12B and 4B models
• DeepSeek V3.2 model
• Kimi K2.5 model
• Inference profile tests for models requiring
• macro in model_verification example to reduce boilerplate

Changed

• BREAKING: Renamed to to match the model's actual version
• Split re-exports into per-vendor blocks to prevent from mixing vendor groupings
• Reordered Claude model definitions by family and version (Opus 4 → 4.1 → 4.5 → 4.6, then Sonnet, then Haiku)
• now uses instead of a hardcoded version

Fixed

• Claude 3.7 Sonnet output token limit: 8,192 → 64,000
• Claude Haiku 4.5 output token limit: 8,192 → 64,000
• Claude Opus 4.5 output token limit: 32,000 → 64,000
• Stale CLI flag in model_verification example docs (correct flag is )