MiniCPM-V-4.7-1B-DSpark-gguf Text Generation Model: 4GB+ VRAM

At a Glance
| Item | Value |
|---|---|
| Repository | openbmb/MiniCPM-V-4.7-1B-DSpark-gguf |
| Family guide | MiniCPM-V-4.7-DSpark guide (2 articles) |
| Publisher guide | OpenBMB: models and licenses |
| Published | 2026-10-15 |
| License | apache-2.0 |
| Formats | GGUF |
| Paper | arXiv:2607.05147 |
| Source type | Primary source (the publisher itself) |
Values determined by this site’s code when the information was collected. Dates are JST.
Overview
Openbmb has released the GGUF version of “MiniCPM-V-4.7-1B-DSpark", a draft model designed to significantly accelerate inference for the multimodal LLM (VLM) “MiniCPM-V-4.7-1B". This model is a checkpoint dedicated to speculative decoding, and when used in combination with the target MiniCPM-V-4.7-1B, it can improve generation speed while maintaining output quality.
“DSpark", adopted by this model, is a new speculative decoding framework proposed in the paper “DSpark: Confidence-Scheduled Speculative Decoding with Semi-Autoregressive Generation". While conventional parallel draft models can propose many tokens at once, they face the issue that acceptance rates drop sharply due to a lack of dependencies between tokens. DSpark introduces a semi-autoregressive architecture, combining lightweight sequential modules to model dependencies within blocks. This aims to achieve both high throughput and high acceptance rates. In addition, during the verification process, it employs “confidence-scheduled verification", which dynamically adjusts the verification length based on prefix survival probability, optimizing overall system efficiency. According to the paper, when introduced into DeepSeek-V4’s serving system, it increased the generation speed per user by 60 percent to 85 percent compared to the conventional baseline.
Specifications
This model has a special configuration as a draft model, and the following specifications have been published:
- Target model: MiniCPM-V-4.7-1B
- Number of draft layers: 5
- Number of draft parameters: 213,547,777
- Number of draft tokens (per forward pass): 15
- Target hidden state layers: [3, 7, 11, 15, 19]
- Checkpoint precision: BF16
- Training sequences: 2M
- Number of epochs: 6
- Optimizer: AdamW
- Objective function: CE + L1 + confidence loss
Performance
In the model card, “Accept Length" is reported as an index to measure the efficiency of speculative decoding. Accept length indicates the average number of tokens accepted (adopted) in a single verification step, and higher values yield greater acceleration benefits. The evaluation used natural EOS (end-of-sentence) stopping and a setting of max_new_tokens=8192.
| Domain | Accept Length |
|---|---|
| Vision | |
| MMMU | 3.452 |
| MathVista | 4.333 |
| CharXiv_reasoning_val | 4.452 |
| OCRBench | 4.752 |
| Text | |
| Math | 5.891 |
| Code | 5.347 |
| General | 3.370 |
According to the measurement results from the publisher, it records high accept lengths of 5.891 and 5.347 particularly in text generation fields such as “Math" and “Code", respectively. This indicates that predictions by the draft model function very efficiently in these tasks with logical structures. Significant acceleration can be expected in coding tasks where functions are implemented according to specifications and in mathematical reasoning.
On the other hand, in the “Vision" field involving image understanding, OCRBench has the highest value at 4.752, while MMMU (college-level problems including charts and photos) remains at 3.452. Compared to OCR tasks that read text in images, tasks requiring advanced image understanding tend to have fewer accepted tokens per step. In addition, the text “General" task is 3.370, showing lower efficiency than specialized math and code tasks.
Note that these figures indicate efficiency as a draft model, and the accuracy of the final generated responses complies with the performance of the target model performing the verification (MiniCPM-V-4.7-1B). When using quantization (the GGUF version), the accept length may fluctuate slightly compared to the original BF16 version, but the content of the generated text itself does not degrade.
Strengths and Use Cases
This model is not a standalone generation model, but a dedicated draft model to accelerate inference for MiniCPM-V-4.7-1B. Therefore, its “strength" lies in bringing out the capabilities possessed by the target model MiniCPM-V-4.7-1B (image understanding, tool calling, lightweight operation for on-device/edge AI, English and Chinese support) without lowering generation speed.
From the tags, vision-understanding, tool-calling, and on-device/edge-ai can be confirmed as features on the MiniCPM-V-4.7-1B side. As seen in the performance section, structural tasks such as math and code have high accept lengths, making them areas where the effect of speculative decoding is prominent. Conversely, tasks that require deep image understanding have slightly lower accept lengths, making the benefits of acceleration relatively smaller.
For use cases, it is suited for situations where you want to run MiniCPM-V-4.7-1B as a chatbot or multimodal assistant in edge devices or local environments while improving response speed. Note that it is not a model to be used alone for chatting.
Hardware Requirements
Estimated requirements (calculated by Local Model Watch) — 214M parameters (taken from the base model openbmb/MiniCPM-V-4.7-1B-DSpark)
| Your VRAM | Quantization | File size | Est. memory needed |
|---|---|---|---|
| 4GB (laptop iGPU / phone class) | F16 | 0.4GB | 0.5GB |
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.
Can You Run It Locally?
Runs in Ollama, LM Studio and llama.cpp as-is.
It is distributed in GGUF, so no conversion is needed.
License — apache-2.0 (Commercial use allowed): Permits commercial use, modification and redistribution. Redistribution requires including the license and stating changes; includes a patent grant.
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. License summaries are not legal advice — check the publisher’s original terms before relying on them.
Recent Models in the Same Size Class
Models with up to 4B 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 |
|---|---|---|---|---|
| openbmb/MiniCPM-V-4.7-35B-A3B-DSpark-gguf | 389M | 4GB | apache-2.0 | MiniCPM-V-4.7-35B-A3B-DSpark-gguf Draft Model for Speedup: 4GB+ VRAM (2026-11-17) |
| IFM/Jais-3-0.9B-GGUF | 1.1B | 4GB | apache-2.0 | Jais-3-0.9B-GGUF Bilingual Dense Model: 4GB+ VRAM (2026-11-17) |
| nvidia/NVIDIA-Nemotron-Labs-Dory1.0-4B-BF16 | 4.0B | 4GB | openmdw-1.1 | NVIDIA Releases Dory1.0-4B: A Recurrent 4B LLM with Dynamic Loops (2026-11-17) |
| LiquidAI/LFM2.5-24B-A2B-DSpark | 328M | 4GB | lfm1.0 | LFM2.5-24B-A2B-DSpark Speculative Decoding Draft Model: 4GB+ VRAM (2026-11-17) |
| LiquidAI/LFM2.5-24B-A2B-DSpark-GGUF | 328M | 4GB | lfm1.0 | LFM2.5-24B-A2B-DSpark-GGUF Text Generation Model: 4GB+ VRAM (2026-11-17) |
How to Get It
This repository is the GGUF (llama.cpp) quantized version of MiniCPM-V-4.7-1B-DSpark. The distribution format is GGUF, the license is apache-2.0, and a gate (agreement to terms of use) is not required.
When using it with llama.cpp, you need to download the target model body, the vision projector (mmproj), and the draft from this repository separately.
huggingface-cli download openbmb/MiniCPM-V-4.7-1B-gguf MiniCPM-V-4_7-1B-Q4_K_M.gguf mmproj-model-f16.gguf --local-dir./1B-gguf
huggingface-cli download openbmb/MiniCPM-V-4.7-1B-DSpark-gguf dspark-MiniCPM-V-4_7-1B-Q8_0.gguf --local-dir./1B-gguf
Execution example (prompt with image, speculative decoding enabled):
llama-cli \
-m./1B-gguf/MiniCPM-V-4_7-1B-Q4_K_M.gguf \
--mmproj./1B-gguf/mmproj-model-f16.gguf \
-md./1B-gguf/dspark-MiniCPM-V-4_7-1B-Q8_0.gguf \
--spec-type draft-dspark \
--image image.jpg -p "Describe this image." \
-ngl 99 -ngld 99 -fa on -c 8192
According to the card, a llama.cpp build supporting the dflash architecture and DSpark Markov head is required. Since the quantization on the draft side does not affect the output content itself and only affects the accept length (speed), Q8_0 or Q6_K are recommended as default values. It is also explained that it can be used in vLLM by specifying "method":"dspark" in --speculative-config.
Related Articles
- MiniCPM-V-4.7-35B-A3B-DSpark-gguf Draft Model for Speedup: 4GB+ VRAM
- OpenBMB Releases Quantized Versions of MiniCPM-OCR-Jev
- MiniCPM-OCR-1.0-DSpark OCR Acceleration Draft Model: 4GB+ VRAM
- MiniCPM5-2B Text Generation Model: Our Test Answers, 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 → MiniCPM-V-4.7-DSpark family overview (2 articles)
- Engines that run this model → llama.cpp / Ollama
- Formats this model is available in → GGUF format guide and models
- Learn about the publisher → OpenBMB: models, licenses and articles

