swama

Local inference engine

aknowledgements: Trans-N-ai/swama

Build Remarks

• Swama/CLI/Run.swift

Task.detached causes compilation error

@MainActor // add
func startAnimation() {
  animationDisplayTask = Task/*.detached*/ {
  ...
  }
}

Convert Model

mlx_lm.convert \
	--hf-path Qwen/Qwen3-4B-Instruct-2507   \
	--mlx-path ~/Desktop/Qwen3-4B-Instruct-2507-4bit  \
	--quantize \
	--q-bits 4 \
	--q-group-size 64

Remove from config.json

{
  "mode": "affine"
}

Only f16 or 4bit is compatible with swama; not 8bit.

Tool calling does not seem to be supported.