Northstar-CUA-Fast-1.6-MLX-4bit

Northstar CUA Fast 1.6 — MLX 4-bit

MLX 4-bit conversion of the Northstar 1.6 line for on-device computer use in Lightcone.

Conversion

• MLX affine 4-bit, 5.35 bits per weight; vision tower retained at source precision
• chat_template.jinja is ALSO embedded as chat_template in tokenizer_config.json — swift-transformers does not read the standalone jinja file, and without the template the model never emits a tool call
• Loader note: mlx-vlm ≤0.6.4 double-applies this architecture's RMSNorm weight shift when loading already-converted artifacts (garbage output). Use mlx-vlm >0.6.4 or a guarded sanitize (should_shift_norm_weights). mlx-swift-lm is safe (metadata format=="mlx" guard).

Action protocol

training_xml / computer_use — the same contract the hosted northstar-cua-fast-1.6 speaks. Prompt with the computer_use function schema in the system prompt; parse exactly one:

<tool_call>
<function=computer_use>
<parameter=action>
left_click
</parameter>
<parameter=coordinate>
[500, 500]
</parameter>
</function>
</tool_call>

Coordinates are integers on a top-left-origin 0–1000 grid. Use deterministic decoding. Do NOT force JSON output — it coerces the model off its trained distribution and clicks drift.