Image tokens not included when running Qwen3.6-35B-A3B MLX (tokens: 0, features N)

LM Studio version: 0.4.12 (Build 1)
Model: unsloth/Qwen3.6-35B-A3B-UD-MLX-4bit
OS: macOS (Apple Silicon)

Error

ValueError: Image features and image tokens do not match: tokens: 0, features 64

Occurs on both the native /api/v1/chat endpoint and the OpenAI-compatible /v1/chat/completions endpoint.

Repro

curl -s http://localhost:1234/api/v1/chat   -H "Content-Type: application/json"   -d '{
    "model": "qwen3.6-35b-a3b-ud-mlx",
    "input": [
      {"type": "text", "content": "What color is this square? One word."},
      {"type": "image", "data_url": "data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAoAAAAKCAYAAACNMs+9AAAAFUlEQVR42mP8z8BQz0AEYBxVSF+FABJADveWkH6oAAAAAElFTkSuQmCC"}
    ],
    "temperature": 0
  }'

Context

This appears to be the same heuristic failure documented in this issue for Qwen3-VL and later Qwen 3.5-35b-a3b (fixed in 0.4.6). The Qwen 3.6 architecture (early-fused multimodal) seems to be triggering the same broken path — the vision encoder runs and produces features, but no image tokens are injected into the prompt.

Start with the Qwen3.6 multimodal request path for both /api/v1/chat and /v1/chat/completions, using the supplied curl request to reproduce the tokens: 0, features 64 mismatch. Compare that behavior with the Qwen3-VL and Qwen 3.5-35b-a3b heuristic failure referenced in the issue. Done means the image tokens and image features match for the reproduced request on both endpoints.