SingVL_S1_4B vs MiniCPM-V 4.7 1B: 4.54 Billion Parameters and No Licence

The most useful number in this matchup is not on either model card. It is 1.55 billion: the tensor total of MiniCPM-V 4.7 1B, the on-device vision model OpenBMB released on October 12, 2026 under Apache 2.0, with GGUF, AWQ, GPTQ, int4 and FP8 builds already published and ports to iOS, Android and HarmonyOS. Against it sits SingVL_S1_4B, inclusionAI's 4.54-billion-parameter decision model, uploaded on November 11, 2026 with no announcement, no licence file and no quantised variant of any kind. Roughly three times the weights, on the wrong side of a phone's memory budget, doing a job that mostly is not the same job. The comparison is worth making anyway, because it exposes the one question a spec sheet cannot answer: what does a model owe you beyond its output?

What each model returns, and therefore where it can live

MiniCPM-V 4.7 1B is an instruct-tuned multimodal model in the conventional sense. It reads a single image, several images, or video, and it generates an answer — captions, transcriptions, visual question answering, multi-turn conversation. OpenBMB ships dedicated DSpark speculative-decoding heads for it, released three days after the model, and supports it across SGLang, vLLM, llama.cpp and Ollama, with fine-tuning through SWIFT and LLaMA-Factory. There is a live demo space and a public cookbook. Every one of those is a channel through which somebody other than the vendor has run the thing.

SingVL_S1_4B does not generate. You hand it a state — text, JSON, up to four images — and a set of named questions with declared types, and it returns a probability for every option you offered. Choice picks a candidate, noul returns the probability that a yes/no proposition holds, and score returns a distribution over two to ten ordered levels plus an expected level. One forward pass, no sampling, no prose. The repository's own quickstart runs a FastAPI server on port 8173 and posts JSON to /v1/systemone.

So the deployment question answers itself before the benchmark question is asked. MiniCPM-V 4.7 1B has an ONNX-adjacent story, phone ports and chip-vendor support because it was designed to sit on a device. SingVL_S1_4B has a Python server because it was designed to sit on a GPU somebody owns. If your product ships an app, one of these is a component and the other is a service dependency.

Efficiency is the whole thesis on one side

OpenBMB's 1B variant exists to hold a position on a performance-versus-inference-time curve, and the card is explicit that this — not raw accuracy — is the claim. The architecture buys that position two ways: intra-ViT early compression, taken from the LLaVA-UHD v4 line, which trims visual tokens inside the vision encoder so later layers do less work, and mixed 4x/16x visual token compression, which lets a caller trade detail for cost per task. On top of that sit the DSpark speculative heads and a stated day-zero support list that runs from Rockchip and Houmo to Intel, AMD and Arm.

The comparison the vendor actually publishes for this variant is against Qwen3.5 0.8B: better overall score and better inference efficiency in Instruct mode, by the vendor's own measurement. That is a narrow claim from the same lab that built the model, and it is the claim the product is built around.

SingVL_S1_4B makes no efficiency claim at all. The card does not report latency, throughput, memory footprint or tokens per second. What it does report, indirectly, is a design that should be efficient at a different task: one forward pass returns every answer to every question in the request, where a generative model needs one pass per token of output. For a five-option classification that is a large structural advantage. For a caption it is not a comparison, because there is no caption.

Worth flagging: the config pins the input contract to "all-input-before-first-answer; single-turn; no-thinking", and the README states 65,536 input tokens for text or 10,000 with images. A 1B model that runs on a phone and a 4.5B model that needs a 10,000-token image budget are solving for different constraints. Neither number is wrong; they are answers to different questions.

The licensing gap is the widest part of this comparison

MiniCPM-V 4.7 1B is Apache 2.0. That is not a marketing detail — it is the reason the ecosystem around it exists. Within four weeks of release there were community GGUF quants from multiple publishers, an MNN port, a Rockchip RKLLM build, an NPU build, a ComfyUI integration and at least one fine-tune on the Hub. None of that could have happened under an unclear licence, because none of those people could have redistributed weights.

SingVL_S1_4B ships with no licence field in its model-card metadata and no LICENSE file in the repository. Under default copyright, that means nobody outside inclusionAI has permission to redistribute it, quantise it and publish the result, or ship it inside a product without a separate agreement. The organisation has published plenty of models with explicit terms — LemonDecider carries Apache 2.0, the SingProbe family carries Q​wen's community licence — so the omission reads as an unfinished release rather than a deliberate restriction. But a reading is not a grant.

The practical consequence: for anything that touches a customer, a procurement process or an app store, MiniCPM-V 4.7 1B is usable this week and SingVL_S1_4B is not. That single line decides more real deployment decisions than every benchmark row on either card.

Benchmarks that must not be put in the same column

There is no shared ruler here, and it is worth being explicit about why rather than quoting numbers at each other.

MiniCPM-V 4.7 1B's evaluation is published as charts — an overall performance comparison, a performance-versus-inference-time frontier — rather than as a per-dataset table in the model card. The numbers exist, the framing is comparative and visual, and the headline claim is positional: better on the frontier than Qwen3.5 0.8B.

SingVL_S1_4B publishes a dense table: 69.76 on the Decision Index Text panel and 73.94 on the Vision panel, with all forty-three text datasets and eleven vision datasets broken out. Those scores come from scoring supplied options and taking the maximum probability, which is a materially easier task than generating a free-form answer for a grader to match, and the Decision Index is explicit that its figures are chance-corrected and coverage-adjusted so guessing earns nothing.

Putting "73.94 vision" next to a MiniCPM-V frontier chart would be a category error dressed as a comparison. The one honest statement is that both sets of figures are vendor-reported, neither has been reproduced by a neutral lab, and SingVL_S1_4B's are additionally two days old with no submission to the Decision Index's public reproduction kit — the card includes the instructions for filing one, which shows intent rather than a result.

Where the 1B model's own numbers get interesting

The Decision Index breakdown for SingVL_S1_4B is more revealing than its aggregate, and it happens to describe a task profile the 1B model also competes for. On document-style vision work, SingVL_S1_4B reports 96.31 on InfographicVQA and 95.82 on the CORD-plus-FUNSD key-information set. On hard academic reasoning it reports 50.17 on MMMU-Pro vision and 49.77 on R-Bench-M. Moderation sits at 71.64 and falls to 43.28 under the skill prompt.

That profile — strong on layout and document extraction, weak on deliberative reasoning, sensitive to prompt framing on subjective judgments — is exactly what you would predict from a 4.5B classifier with a vision encoder and no generation step, and it maps onto the same work MiniCPM-V 4.7 1B is sold for: documents, screens, on-device OCR. So the two models do overlap on task, even though they differ on output shape. The question a team actually faces is whether it needs the extracted value or the extracted answer.

If the pipeline needs the field, the number, the label — SingVL_S1_4B's output is already the thing you wanted, and you skip the parsing. If the pipeline needs the page rendered as text with structure preserved, or an explanation of why the field was ambiguous, MiniCPM-V 4.7 1B is the only one of the two that can produce it.

The middle path, and what it costs

Neither of these has to be an architectural commitment made on day one, and the cheaper experiment is not the obvious one. MiniCPM-V 4.7 1B runs on a laptop, so testing it costs an afternoon and nothing else. SingVL_S1_4B costs a GPU and a 9.1 GB download to test at all, and its licence situation means the test cannot become a deployment without a conversation nobody at inclusionAI is currently having in public.

If what you want to know is whether typed-decision scoring helps your task at all — before you invest in hosting a new checkpoint — the pattern is already callable. OrcaRouter carries TypeSafe Jev 1.14 at $0.042 per million input tokens alongside 200-plus other models on one key, with provider list pricing passed through at 0% markup and automatic failover if a provider drops. That gives you a working decision endpoint to benchmark against your own labels, next to the generative models you would keep for everything else. SingVL_S1_4B is not on OrcaRouter and there is no hosted endpoint for it anywhere, so nothing here is a claim about its availability; the point is that the design question and the vendor question are separable, and only one of them needs a GPU.

The honest summary, in the order the decisions get made

• Can it legally go in a product — MiniCPM-V 4.7 1B, Apache 2.0, today. SingVL_S1_4B, no, until a licence appears.

• Can it run where the product runs — MiniCPM-V 4.7 1B, on phones and edge silicon. SingVL_S1_4B, on a GPU server you operate.

• Does it return what the pipeline needs — SingVL_S1_4B if the answer is a label and a confidence; MiniCPM-V 4.7 1B if it is text, structure or explanation.

• Is either independently proven — no. Both sets of numbers are the vendors'. MiniCPM-V 4.7 1B has four weeks of third-party use behind it; SingVL_S1_4B has two days and zero downloads.

• Which would I evaluate first — MiniCPM-V 4.7 1B, because it costs nothing to try and cannot be ruled out on paperwork. Then the decision pattern against a served endpoint, and only then the new checkpoint, if the pattern pays for itself.

That ordering is not a verdict on the 4.5-billion-parameter model. It is a statement about what a release without a licence, a paper or an announcement can ask of a stranger, which is not very much. The artefact is real and the design is interesting; both of those can be true while the answer to "should we use it" stays no for a while yet.