← All posts

Drafta 1.2: local models, model selection, and calculator

Translated automatically from RussianRead the original →

Drafta 1.2 connects local models via LM Studio and Ollama, lets you choose a model for each provider, adds OpenCode Zen and a built-in calculator.

Updated:

In 1.2, AI in Drafta is no longer tied to a single provider and a single model. The main thing: you can now work with a model on your own Mac, without keys and without sending text to the cloud.

Local models

The new Local provider works with any OpenAI-compatible server: LM Studio, Ollama, llama.cpp. No key is needed, and the text never leaves your computer. The Test connection button shows which models are available on the server.

I measured on an M4 Max with Qwen 3.5:

ModelFirst word of the suggestionGeneration
Qwen3.5-4B, MLX 4-bit0.13–0.16 s~116 tokens/s
Qwen3.5-9B, MLX 4-bit0.22–0.24 s~67 tokens/s

For ghost text, 4B is enough: the suggestion appears almost immediately. 9B is better for editing text and mermaid diagrams via ⌘J.

Your own model for each task

  • Model selection is available for every provider: Anthropic, OpenAI, DeepSeek, Local, OpenCode Zen. The list is pulled from the server, and the name can be entered manually.
  • The assistant and ghost text are configured separately. For example, Claude for editing via ⌘J and a small local model for instant suggestions.
  • Reasoning level for each role: Off, Minimal, Low, Medium, High. The default is Off: fast and no more expensive than before. Drafta itself translates the choice into the parameter for the required provider; GPT, Gemini, Claude, and DeepSeek all have different ones.

OpenCode Zen

The opencode.ai gateway with an inexpensive catalog of models: Claude, GPT, Gemini, DeepSeek, Qwen, GLM, Kimi, and free ones. You paste in the Zen key, and all supported models appear in the list. For each one, Drafta itself selects the required address and request format.

Calculator in the line

You type 1250*12*0,87=, and the result appears immediately. Tab or Enter inserts it.

  • Operations + − × ÷ ^, parentheses, percentages (15% от 2400=, 2400-15%=), sqrt, round, min, max.
  • Decimal comma: however you enter it, that’s how it will output it.
  • In code, formulas, and links, the calculator stays silent.

While you’re typing an expression, ghost text does not make suggestions: only the calculator computes, accurately and without a model. I made this decision after measuring. Qwen 9B laid out a rent calculation step by step correctly, but in the final line it named a number that was not in its own calculation.

Fixed

  • The Accept button in the ⌘J window could insert the response again and again without closing the window. This happened if the response contained a numbered list.
  • Model responses no longer mention the service marker [insert here].

The update will arrive on its own, or download it at drafta.org.