SingVL_S1_4B Multimodal Decision Model with Image Support: 4GB+ VRAM

- 1. At a Glance
- 2. Overview
- 3. Specifications
- 4. Performance
- 5. Strengths and Use Cases
- 6. Hardware Requirements
- 7. Can You Run It Locally?
- 8. Our Own Measurements
- 9. Recent Models in the Same Size Class
- 10. How to Get It
- 11. Quantized and Converted Variants
- 12. Related Articles
- 13. What to Read Next
- 14. Sources
- 15. Update History
At a Glance
| Item | Value |
|---|---|
| Repository | inclusionAI/SingVL_S1_4B |
| Family guide | SingVL-S1 guide (2 articles) |
| Publisher guide | inclusionAI (Ant Group): models and licenses |
| Published | 2026-11-11 |
| Formats | safetensors |
| Source type | Primary source (the publisher itself) |
Values determined by this site’s code when the information was collected. Dates are JST.
Overview
inclusionAI’s inclusionAI/SingVL_S1_4B (Sing VL S1) is a multimodal “decision model" based on Qwen3.5. It answers named questions about text or JSON alongside up to 4 images, scoring the provided choices in a single Transformers forward pass. Rather than generating text in a chat format, it is intended to be used as a component for classification, judgment, and scoring.
It provides 3 types of outputs:
- Choice: Selects a choice and returns probabilities for each candidate
- Noul: Returns the probability that a yes/no proposition is true
- Score: Returns probabilities for 2 to 10 levels along with the expected level
Specifications
- Base: Qwen3.5 (architecture is
Qwen3_5ForConditionalGeneration) - Task format: image-text-to-text
- Input images: Up to 4 per request
- Input limits: 65,536 tokens for text-only per question, or 10,000 tokens when images are included
- Library: transformers (safetensors)
Performance
Here is the comparison table for Public Decision Index 0.3 listed on the model card. The figures are self-reported by the publisher and have not been verified by third parties.
→ 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 |
In this table, Sing VL S1 achieves the highest values among the compared models for both Text and Vision. For Text, it leads the second-place StartLux-Decision v1.1 4B (67.8) by 1.96 points, and holds a gap of around 9 points from the other two models. For Vision, it scores 73.94, a narrow lead of 1.44 points over Perplexity Decider Flash v1.2 (4B) at 72.5. TypeSafe Jev 1.14’s Vision score is marked as “—" and cannot be compared. Note that the Text score of 69.76 matches the overall “Skill" column total in the card’s Public Text Index table (the Non-Skill column in the same table is 77.26). Because the model card excerpts do not define Skill/Non-Skill, the difference cannot be explained.
Next, here is an excerpt of the main items from Public Text (43 datasets):
| Dataset | Original metric | Non-Skill | Skill |
|---|---|---|---|
| BFCL | Case exact accuracy | 96.81 | 95.70 |
| MMLU † | Accuracy | 82.93 | 77.24 |
| GPQA Diamond | Accuracy | 53.06 | 37.41 |
| ARC-Challenge † | Accuracy | 96.25 | 95.00 |
| WinoGrande | Accuracy | 89.82 | 79.64 |
| HellaSwag | Accuracy | 96.48 | 95.31 |
| GSM8K | Accuracy | 73.84 | 68.68 |
| HLE | Accuracy | 12.77 | 0.00 |
| MMLU-Pro | Accuracy | 68.40 | 64.46 |
| BBH | Accuracy | 88.61 | 83.49 |
(† indicates datasets that the Decision Index treats as a separate aggregation and excludes from the Text Index.)
BFCL scores high at 96.81/95.70, making it suitable for tasks that determine tool selection and argument assembly from options. Common sense reasoning benchmarks like HellaSwag and ARC-Challenge also reach high levels above 95, though HellaSwag is a metric that is saturated on larger models. On the other hand, it scores lower on difficult question sets such as GPQA Diamond at 53.06/37.41 and HLE at 12.77/0.00 (though HLE is a metric where even scores in the tens of percent are considered high). GSM8K also stays at 73.84/68.68.
In Vision, InfographicVQA scores high at 96.31/95.08 and KIE (CORD+FUNSD) at 95.82/93.12, whereas R-Bench-M stays at 49.77/39.72 and the Skill column for Moderation (Hateful Memes) is 43.28. The Public Vision Index is 73.94.
Note also that the Non-Skill column is generally higher than the Skill column throughout. How you interpret the model depends on which one you look at.
Strengths and Use Cases
Sing VL S1 is not a model that freely generates text; instead, it is a decision-making model that returns probabilities for pre-defined choices or levels. According to the model card, its use cases are organized into three formats:
- Choice: Classification and selection choosing one from multiple candidates
- Noul: Judging the probability that a yes/no proposition is true
- Score: Scoring across 2 to 10 levels (yielding probabilities for each level and the expected level)
It accepts text, JSON, and up to 4 images as inputs, allowing the same question format to be used with images. The card includes an example of having the model choose “which image is red".
In the benchmark tables (self-reported by the publisher), high scores are observed in BFCL, InfographicVQA, and KIE (CORD+FUNSD). Conversely, scores are lower on challenging reasoning sets like GPQA Diamond and HLE, as well as R-Bench-M. Therefore, while it shows promise as a component for routine judgment, classification, and scoring, it may not be suitable for tasks requiring complex reasoning over difficult problems.
Tags attached include decision-model, typed-decisions, and multimodal. The scope of multilingual support cannot be determined from the documentation.
Hardware Requirements
Estimated requirements (calculated by Local Model Watch) — 4.5B parameters
| Your VRAM | Quantization | File size | Est. memory needed |
|---|---|---|---|
| 4GB (laptop iGPU / phone class) | Q6_K | 3.2GB | 3.9GB |
| 8GB (RTX 4060 / 3060 Ti, etc.) | Q8_0 | 4.2GB | 5.0GB |
| 12GB (RTX 4070 / 3060 12GB, etc.) | F16 | 7.8GB | 9.4GB |
Inference engine support (architecture name matched against each project’s own model registry in its source code, checked 2026-11-17): llama.cpp: registered, vLLM: registered, MLX (mlx-lm): registered.
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. File sizes are measured from the converted build mradermacher/SingVL_S1_4B-GGUF. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.
Can You Run It Locally?
Runs in Ollama, LM Studio and llama.cpp via a converted build.
The publisher ships safetensors, but mradermacher/SingVL_S1_4B-GGUF provides a GGUF build you can use.
Compression: the Q6_K build measures 6.06 bits per weight — about 38% the size of the original 16-bit weights, calculated by this site from the actual file sizes.
Compiled by this site’s code from the published formats, converted builds we have found, and each engine’s own model registry. “Not found" means we have not seen such a build, not that none exists.
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).
Recent Models in the Same Size Class
Models with 4–15B parameters that Local Model Watch covered recently, listed by code from our article log for comparison. VRAM tiers are this site’s estimates; licenses are as stated on the model cards.
→ Scroll horizontally to see all columns
| Model | Parameters | Smallest VRAM tier | License | Our article |
|---|---|---|---|---|
| tencent/Youtu-Parsing-Omni | 5.3B | 12GB | youtu-parsing | Youtu-Parsing-Omni Omnimodal Analysis Model: 12GB+ VRAM (2026-11-17) |
| nvidia/Muse-Spark-Open-816B-A42B-DFlash2 | 4.4B | 12GB | nvidia-open-model-agreement-license | Muse-Spark-Open-816B-A42B-DFlash2 Vision-Language Model: 12GB+ VRAM (2026-11-17) |
| tencent/WeVisOne-VL-4B | 4.8B | 4GB | apache-2.0 | Tencent Releases WeVisOne-VL Vision Language Models (2026-11-17) |
| inclusionAI/ZwZ2-9B-GGUF | 9.4B | 8GB | apache-2.0 | ZwZ2-9B-GGUF Vision-Language Model: 8GB+ VRAM (2026-11-17) |
| inclusionAI/ZwZ2-4B-GGUF | 4.5B | 4GB | apache-2.0 | ZwZ2-4B-GGUF Vision-Language Model: 4GB+ VRAM (2026-11-17) |
How to Get It
- Distribution format: safetensors (for Transformers)
- No license agreement is required to access it (unguided/no gate)
- Supported engines: The card points to the included inference code (Python API using the
SingVLS1class) and an HTTP server via FastAPI (serve.py). Distributions for GGUF or Ollama are not mentioned in the documentation
Enter inference/ from the repository root and install the dependencies. Python 3.12 and PyTorch/torchvision matching your hardware are required.
cd inference
python -m pip install -r requirements.txt
Load it from Python as follows:
from model import SingVLS1
model = SingVLS1("..", device="cuda:0")
To use it as an HTTP API, start the server with the following command:
python serve.py --model.. --port 8173 --image-root./examples
Once started, send a JSON request to POST http://127.0.0.1:8173/v1/systemone. Images can be passed via data_url, base64, or server-side path.
The card also outlines procedures for using the Decision Index kit to reproduce the same evaluations in your own environment and submit the results.
Quantized and Converted Variants
→ Scroll horizontally to see all columns
| Added | Publisher | Format | Repository | Smallest VRAM tier (build, est. memory) |
|---|---|---|---|---|
| 2026-11-17 | mradermacher | GGUF | mradermacher/SingVL_S1_4B-GGUF | Q6_K 3.9GB (fits in 4GB VRAM) |
File sizes of each build:
- Available builds in mradermacher/SingVL_S1_4B-GGUF: Q2_K 1.8GB / Q3_K_S 1.9GB / Q3_K_M 2.1GB / Q3_K_L 2.3GB / IQ4_XS 2.4GB / Q4_K_S 2.4GB / Q4_K_M 2.5GB / Q5_K_S 2.8GB / Q5_K_M 2.9GB / Q6_K 3.2GB / Q8_0 4.2GB / F16 7.8GB
In addition, 1 converted build(s) from other uploaders exist on Hugging Face; this site lists only builds from the model’s publisher or established quantization maintainers.
This section is appended automatically by Local Model Watch when a converted build of this model appears after publication. Memory figures are estimated from the size of the distributed files. What the quantization names mean: glossary.
Related Articles
- SingVL_S1_4B_GGUF 4B Multimodal Decision-Making Model: 4GB+ VRAM
- ZoomGain-4B: 4GB+ VRAM, GGUF Builds
- ZoomGain-9B Fine-Grained Visual Language Model: 8GB+ VRAM, GGUF Builds
- ZwZ2-4B-GGUF Vision-Language Model: 4GB+ VRAM
What to Read Next
- Find models by VRAM (This model runs from the 4GB tier) → Other models that run on a 8GB GPU
- Explore the same model family → SingVL-S1 family overview (2 articles, 1 converted builds)
- Engines that run this model → llama.cpp / Ollama / vLLM
- What Q6_K, Q8_0 mean and where to get this model → GGUF format guide and models / Quantization and model-format glossary
- Learn about the publisher → inclusionAI (Ant Group): models, licenses and articles
Sources
Update History
- 2026-11-17: Added converted builds to “Quantized and Converted Variants”: mradermacher/SingVL_S1_4B-GGUF
- 2026-11-17: Updated the hardware requirements table with the actual file sizes of mradermacher/SingVL_S1_4B-GGUF.
- 2026-11-17: Changed the title to show what the article covers (VRAM requirements, file list, etc.).

