Thinking effort and instant dictation

Hey everyone,

We are releasing two quality-of-life improvements to Alter focusing on a core goal: giving you more precise control over the trade-off between reasoning depth and execution speed.

🎶 Kano - Ikeya Seki 🎶 (Thank you for the ride Kavinsky)

Optimizing Model Thinking Effort

Not every task requires the same level of cognitive "heavy lifting." Using a maximum-reasoning model for a simple grammar fix is often a waste of time and latency.

You can now manage the Thinking Effort of your models directly from the model picker (accessed via /). We have simplified this into three clear options:

1. Low: Optimized for speed-sensitive, non-complex tasks. Ideal for correcting grammar, cleaning up dictation, or basic formatting.

2. Balance: The default setting, suitable for the majority of general-purpose interactions.

3. Max: Full reasoning power for complex logic, deep analysis, and difficult problem-solving.

To switch levels, simply use the left and right arrow keys in the model picker or click the desired effort level with your mouse.

Instant Dictation

Dictation is a primary way many of you interact with Alter, but the gap between speaking and seeing text on screen has always been a point of friction.

We have implemented a entirely new dictation pipelinethat is significantly faster than before. When combined with the ability to disable dictation cleanup, the results are nearly instantaneous. Your transcribed text now appears on the screen in real-time, creating a much more fluid and responsive experience.

Deepseek V4 has been upgraded

You can now use Deepseek-v4-Flash-0731, a highly efficient and powerful model from Deepseek.

Cheers,
Olivier & Sam

Full Changelog

New Features & Enhancements

• Thinking effort: Choose each model's advertised thinking effort from Hub, QuickHub, or the Action Editor. Alter remembers the choice per model and uses it for future chats and actions.

• Dictation: Improve dictation speed for parakeet processor.

• Custom Endpoints: Context Window size, thinking and vision capabilities are detected automatically for LM Studio, OpenRouter, Ollama, Bifrost and Gemini.

• Action Editor: A new searchable model picker groups the current model, favorites, and other models, with controls for thinking effort, context, image support, and favorites.

• Dictation shortcuts: Dictation now accepts modifier-only shortcuts, including Command, left or right Control, and left or right Option.

• Discord sources: Web search sources are formatted more clearly in Discord responses.

• New models: New models: GLM-5.2 and deepseek v4 flash 0731

Bug Fixes & Stability

• Model selection: Recover automatically when a saved model is no longer available, align switcher metadata consistently, and clarify vision support without redundant selection indicators.

• Tools: Code Runner now shows its sandbox tool.