Changelog
What's new in Refrase.
Fix
v0.1.3 — Fixed Qwen thinking mode detection
- Fixed incorrect /think vs /no_think detection for Qwen3 models in analysis tasks
- Thinking mode is now properly disabled for extraction tasks to avoid unnecessary latency
- Added test coverage for all Qwen3 task-type and thinking-mode combinations
Improvement
v0.1.2 — Improved XML structuring for Claude
- Smarter XML tag placement preserves inline formatting and lists
- Role extraction now handles multi-sentence persona definitions
- Output format guidance tags are appended after existing instructions instead of replacing them
New Model
v0.1.1 — Added 11 model families
- Claude (Sonnet 4.6, Opus 4.6, Haiku 4.5)
- GPT (4o, 4o-mini, o3, o4-mini)
- Gemini (Pro, Flash, Ultra)
- Qwen3, DeepSeek V3, Mistral Large 3, Llama 3.1
- Kimi K2, GLM 4.7, Nemotron 30B, MiniMax M2
- 39+ model variants across all families
Feature
v0.1.0 — Launch: Refrase v1.0
- 11 model family adapters (39+ variants)
- Python and TypeScript SDKs published to PyPI and npm
- MCP server for Claude Desktop and Cursor
- Web app with /adapt tool and /build chatbot
- Browser extension for Chrome and Firefox
- Dashboard with saved prompts and usage tracking
- Model comparison pages with side-by-side benchmarks