SingVL S1 4B: a multimodal decision-making model built on Qwen3.5

On November 11, 2026, the model inclusionAI/SingVL_S1_4B was published on HuggingFace with the image-text-to-text task. It is a multimodal decision-making model built on Qwen3.5: it answers questions based on text or JSON and accepts up to four images. Each question is processed in a single Transformers pass that scores the proposed options. The model supports three modes: Choice selects an option and returns a probability for each candidate, Noul estimates the probability that a "yes/no" statement is true, Score rates the input on a scale of 2 to 10 levels and returns the probability of each level and the expected level.

What it means

The model has 4.5 billion parameters. The largest weights file is 4.7 GB, and the repository totals 8.5 GB in safetensors format. It runs via transformers. The site hosts two quantized GGUF builds, the smallest working one being 2.3 GB in IQ4_XS format from maternion. The author did not specify a license. Running the full model locally will require a GPU with enough VRAM for 8.5 GB of weights; the 2.3 GB quantized build noticeably lowers the requirements. According to the benchmarks in the model page, Sing VL S1 scores 69.76 on the text index and 73.94 on the visual one, beating TypeSafe Jev 1.14, StartLux-Decision v1.1 4B and Perplexity Decider Flash v1.2 on text tasks. For more news on local AI tools, see the CPU3D news feed.