Download Latest Version
v2.1.0 source code.tar.gz (147.2 MB)
Get an email when there's a new version of LLM.swift
| Name | Modified | Size | Downloads / Week |
|---|---|---|---|
| Parent folder | |||
| README.md | 2025-12-06 | 981 Bytes | |
| v2.1.0 source code.tar.gz | 2025-12-06 | 147.2 MB | |
| v2.1.0 source code.zip | 2025-12-06 | 147.4 MB | |
| Totals: 3 Items | 294.6 MB | 0 | |
v2.1.0
Thinking Output Separation
You can now access the model's reasoning process separately from the final response.
Usage
:::swift
await bot.respond(to: "What is 2+2?", thinking: .enabled)
print(bot.thinking) // Model's reasoning process
print(bot.output) // Final answer
ThinkingMode
| Mode | Behavior |
|---|---|
.none |
Default. No special handling. |
.enabled |
Streams thinking and response separately. |
.suppressed |
Injects empty thinking tokens to skip reasoning. |
Compatibility
This feature works best with models that have native thinking token support (e.g., Qwen, DeepSeek-R1).
For models without dedicated thinking tokens, the library uses string-based marker detection which may still work, but performance improvements are not guaranteed. Experimentation is recommended.
Full Changelog: https://github.com/eastriverlee/LLM.swift/compare/v2.0.2...v2.1.0
