SingVL_S1_4B_GGUF 4B Multimodal Decision-Making Model: 4GB+ VRAM

SingVL_S1_4B_GGUF 4B Multimodal Decision-Making Model: 4GB+ VRAM

At a Glance

Item Value
Repository inclusionAI/SingVL_S1_4B_GGUF
Family guide SingVL-S1 guide (2 articles)
Publisher guide inclusionAI (Ant Group): models and licenses
Published 2026-11-12
Formats GGUF
Source type Primary source (the publisher itself)

Values determined by this site’s code when the information was collected. Dates are JST.

Overview

inclusionAI has released the GGUF version repository “inclusionAI/SingVL_S1_4B_GGUF" for the 4B parameter multimodal decision model “Sing VL S1".

Built on Qwen3.5, this model accepts text, JSON data, and up to four images as input, scoring the provided choices in a single forward pass.

Three specific types of task formats are available:
– Choice: Selects an option and returns the probability for each candidate.
– Noul: Returns the probability that a yes/no proposition is true.
– Score: Evaluates the input on a 2 to 10 level ordered scale from low to high, returning the probability for each level and the expected value calculated from the sum of each level index multiplied by its probability.

Internally, the full attention layers use bidirectional attention for the full input, while the GDN layers maintain causal recurrence. At the start of each question, the KV cache and recurrence state are cleared, and candidate logits are read from the final valid position of the input.

Specifications

  • Parameters: 4B
  • Architecture: Qwen3.5
  • Context Length: 4096 tokens

Performance

Comparison results of the public reference score “Public Decision Index 0.3" for the original model “Sing VL S1" listed on the model card are as follows. This test is based on all public runs of the original model (140,178 text requests and 12,210 vision requests).

→ Scroll horizontally to see all columns

Panel Sing VL S1 TypeSafe Jev 1.14 StartLux-Decision v1.1 4B Perplexity Decider Flash v1.2 (4B)
Text 69.76 60.70 67.8 60.1
Vision 73.94 — 67.90 72.5

From the published comparison table, the original model Sing VL S1 achieved 69.76 in the text panel and 73.94 in the vision panel, recording the highest level of scores in both categories among the compared models.

In particular, for text evaluation, it showed superior results with a margin of several points or more compared to other comparison models (60.1 to 67.8). On the other hand, for vision evaluation, the difference was 1.44 points compared to 72.5 recorded by Perplexity Decider Flash v1.2 (4B), which is a slightly smaller margin compared to the text domain.

Note that all the figures above are measurement results for the original model (Sing VL S1). Verification as a GGUF variant is currently reported to cover 9 examples and 11 questions (across BF16, Q8_0, and Q4_K_M quantizations), with complete public score measurements for the GGUF version pending.

Strengths and Use Cases

This model excels at decision-making tasks where it analyzes text, JSON data, and up to four image inputs to output judgments for type-defined structured questions.

Specifically, based on the model card and capability tags, it is suitable for the following applications:

  • Choice Tasks: Selects the optimal choice from defined options with probabilities based on the given text or image content. It is suited for image recognition and classification judgments.
  • Binary Judgment Tasks (Noul): Objectively evaluates binary propositions such as “whether specified conditions are met" by calculating the probability of being true.
  • Multi-level Score Evaluation (Score): Evaluates the input content using predefined levels from 2 to 10, calculating the probability distribution and expected value for each level. It is used for text and image quality assessment or classification strength determination.

Additionally, because it supports up to four image inputs, it is suitable for comparative judgments targeting multiple images and complex multimodal analysis combining images and text. These strengths and use cases are based on the capabilities of the original model, “Sing VL S1".

Hardware Requirements

Estimated requirements (calculated by Local Model Watch)

Your VRAM Quantization File size Est. memory needed
4GB (laptop iGPU / phone class) Q4_K_M 2.5GB 3.0GB
8GB (RTX 4060 / 3060 Ti, etc.) Q8_0 4.2GB 5.0GB
12GB (RTX 4070 / 3060 12GB, etc.) BF16 7.8GB 9.4GB

Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model’s authors. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.

Our Own Measurements

Measurements We Did Not Take

This model does not state a license, so we cannot confirm permission to use it. Because this site carries advertising, we did not run the model (no CPU run, answers, quantization comparison, conversion or generation).

How to Get It

This model is released in GGUF format, providing decoders (BF16, Q8_0, Q4_K_M) along with a common vision projector (mmproj-SingVL-4B-F16.gguf). Gated license agreement procedures are not required, allowing it to be acquired and used as-is.

For the hardware environment, a dedicated adapter built with llama.cpp as the backend is used in a Linux x86_64 CPU environment. Python 3.12 and corresponding PyTorch and torchvision setups are required for usage.

Model and vision projector paths, CPU thread counts, and context lengths are configured via environment variables, and inference is executed via Python scripts or HTTP API services.

A usage example from Python code is as follows:

import os
from PIL import Image
from examples.llama_cpp.model import SingVLS1

os.environ["JEV_GGUF_MODEL"] = "models/SingVL-4B-Q8_0.gguf"
os.environ["JEV_GGUF_MMPROJ"] = "models/mmproj-SingVL-4B-F16.gguf"
os.environ["JEV_CPU_CONTEXT"] = "4096"
os.environ["JEV_CPU_THREADS"] = "8"

model = SingVLS1("config", device="cpu")
try:
    text, _ = model.predict(
        state="The color is red. ",
        questions={
            "color": {
                "type": "choice",
                "instructions": "Which color is named?",
                "criteria": {"red": "Red", "blue": "Blue"},
            },
            "yesno": {"type": "noul", "instructions": "Is the named color red?"},
            "redness": {
                "type": "score",
                "instructions": "How clearly does the text identify the color as red?",
                "criteria": ["Not red", "Possibly red", "Clearly red"],
            },
        },
    )
    print(text["answers"])
finally:
    model.close()

When starting as an HTTP API server using FastAPI, prepare the adapter according to the repository’s build instructions and then launch the server with the following command:

export JEV_GGUF_MODEL="$PWD/models/SingVL-4B-Q8_0.gguf"
export JEV_GGUF_MMPROJ="$PWD/models/mmproj-SingVL-4B-F16.gguf"
export JEV_CPU_CONTEXT=4096
export JEV_CPU_THREADS=8
export OMP_NUM_THREADS=8

python examples/llama_cpp/serve.py --model "$PWD/config" \
  --device cpu --port 8173 --image-root "$PWD/examples/images"

After starting the server, inference results can be obtained by sending a JSON-formatted request to the /v1/systemone endpoint using a curl command like this:

curl -fsS http://127.0.0.1:8173/v1/systemone \
  -H 'Content-Type: application/json' -d '{
  "model": "sing-vl-s1",
  "state": "The color is red. ",
  "questions": {
    "color": {
      "type": "choice",
      "instructions": "Which color is named?",
      "criteria": {"red": "Red", "blue": "Blue"}
    },
    "yesno": {"type": "noul", "instructions": "Is the named color red?"},
    "redness": {
      "type": "score",
      "instructions": "How clearly does the text identify the color as red?",
      "criteria": ["Not red", "Possibly red", "Clearly red"]
    }
  }
}'

When using images, specify Base64 string data (data_url or base64) or an image path on the server (path) in the images array.

Related Articles

What to Read Next

Sources