A new generation of "small but mighty" arrives! 1.3B model wins on both performance and efficiency as MiniCPM-V 4.6 goes open source

2026-05-13

On May 11, ModelBest, together with Tsinghua University and the OpenBMB open-source community, officially released and open-sourced a new generation of on-device multimodal large model: MiniCPM-V 4.6.

It is the best-balanced model between efficiency and performance in MiniCPM-V family history. With only 1.3B parameters, it achieves a double breakthrough in performance and efficiency, tops models of the same size globally, and comprehensively surpasses Alibaba's Qwen3.5-0.8B and Google's Gemma4-E2B-it.

Especially at a time when memory prices are soaring, MiniCPM-V 4.6 runs smoothly on-device with only 6 GB of memory, truly achieving "low memory, blazing speed." This not only greatly lowers the barrier to deploying multimodal large models on all kinds of smart devices, but also brings ModelBest's "AGI for Lives" vision one step closer.

This article includes video content. Please view it in the original WeChat post.

MiniCPM-V 4.6 is now fully open-sourced on GitHub, Hugging Face, and other platforms, and a TestFlight beta is available for phones running iOS, Android, and HarmonyOS.

➤ Hugging Face

🔗https://huggingface.co/openbmb/MiniCPM-V-4.6

➤ GitHub

🔗 https://github.com/OpenBMB/MiniCPM-V

➤ ModelScope

🔗 https://modelscope.cn/models/OpenBMB/MiniCPM-V-4.6

➤ GitCode

🔗https://ai.gitcode.com/OpenBMB/MiniCPM-V-4.6

➤ Web Demo

🔗 https://huggingface.co/spaces/openbmb/MiniCPM-V-4.6-Demo

➤ APP Demo

🔗 https://github.com/OpenBMB/MiniCPM-V-Apps

Outstanding performance: 1.3B delivers best-in-class results

According to multiple authoritative benchmark evaluations, both versions of MiniCPM-V 4.6 (Instruct and Thinking) deliver excellent overall multimodal capability, leading across every dimension.

Instruct version: comprehensively surpasses models such as Qwen3.5-0.8B and Gemma4-E2B-it on general image-text understanding, STEM mathematical reasoning, document OCR, video temporal understanding, and object grounding.

Thinking version: also leads across almost all high-level tasks such as multi-image relational reasoning and hallucination suppression.

On the latest Artificial Analysis (AA) leaderboard, MiniCPM-V 4.6 ranks among the top with an excellent score of 13, surpassing a range of models including Mistral 3-3B and Qwen 3.5-0.8B and approaching Qwen 3.5-2B, making it the undisputed performance benchmark among 1B-class open-source models.

Efficiency overtakes: redefining "intelligence density"

Beyond its striking performance, MiniCPM-V 4.6 achieves an efficiency breakthrough that could be called "counterintuitive."

Although it has more parameters than Qwen3.5-0.8B, MiniCPM-V 4.6 runs faster, achieving a striking overtake:

Inference throughput: vLLM-based token throughput is 1.5× that of Qwen3.5-0.8B.

Compute cost: in the AA evaluation, it surpasses Qwen3.5-0.8B using only 2.5% of the token consumption (5.4M vs 233M).

This means MiniCPM-V 4.6 leverages an extremely small compute cost to achieve an extremely high level of intelligence, another embodiment of ModelBest's "Density Law."

As a result, with the same hardware, developers can serve several times the online traffic or achieve even faster response on-device.

This leap means that almost every personal device on the market (phones, computers) can run MiniCPM-V 4.6 efficiently.

The hardcore innovation behind the extremes

Behind the striking performance and efficiency are hardcore innovations in two underlying technologies.

ViT architecture rework: image-encoding compute cut by 50%

The vision encoder (ViT) is the core component through which multimodal models process image information, and one of the most compute-intensive stages.

MiniCPM-V 4.6 adopts the latest LLaVA-UHD v4 technology, jointly developed by ModelBest and Tsinghua University. Through an innovative "early visual token compression inside the ViT," it greatly reduces the number of visual tokens and the amount of computation without sacrificing performance, cutting image-encoding overhead by 50% compared with a conventional ViT.

Addressing the efficiency of visual encoding when models process high-resolution images, LLaVA-UHD v4 proposes an efficient encoding paradigm that reduces floating-point operations in the visual encoding stage by 55.8% while surpassing the original performance.

LLaVA-UHD v4 paper: https://huggingface.co/papers/2605.08985

LLaVA-UHD v4 has two core designs:

Efficient slice encoding: replacing conventional global encoding, it splits a high-resolution image into multiple blocks for processing, structurally avoiding the bottleneck of attention compute growing quadratically with resolution.

Early compression inside the ViT: a compression module is introduced in the shallow layers of the ViT, so that most subsequent ViT layers process only a very small number of tokens, reducing computation at the root.

This rework works in concert with 16× token compression to enable extremely lightweight inference — completing image understanding tasks of equal quality with fewer compute resources. It also creates the efficiency miracle of MiniCPM-V 4.6: "more parameters, yet faster."

4×/16× hybrid token compression: performance or speed, switch on demand

Industry multimodal foundation models generally use a fixed 4× visual token compression rate, but MiniCPM-V 4.6 breaks with convention and offers an industry-leading dual-mode switch:

4× compression mode (performance first): suited to demanding document parsing, dense text recognition, and other detail-sensitive tasks;

16× compression mode (speed first): suited to real-time interaction, low-compute environments, high-concurrency deployment, and similar scenarios.

One model, two options — developers no longer have to make a painful trade-off between performance and speed.

The visual token compression rate affects many key efficiency metrics, including VRAM usage, first-response latency, inference throughput, and power consumption; the higher the compression rate, the faster the response. MiniCPM-V multimodal models introduced a 16× compression rate back in 2024, leading the industry.

Notably, ModelBest's pioneering 16× compression technology has long been validated in industry. For example, Kuaishou's OneRec recommendation model, released in 2025, uses the MiniCPM-V series when processing massive video multimodal features, successfully supporting 25% of the traffic requests in its main scenario.

Thanks to 16× visual token compression, MiniCPM-V 4.6 establishes significant advantages in two key dimensions of inference efficiency at once:

For single-concurrency time to first token (TTFT), it flattens the "resolution-latency" curve almost completely: a 3136² high-definition image has a TTFT of only 75.7 ms, 2.2× faster than Qwen3.5-0.8B of the same base scale;

For high-concurrency throughput, a single GPU reaches 7013 tokens/s and 54.79 images/s for 1344² images, 1.5× that of Qwen3.5-0.8B, meaning the same hardware can serve several times the online traffic.

The metrics in both dimensions point to the same conclusion:

With shorter visual sequences and a smaller KV-Cache, MiniCPM-V 4.6 pushes both the on-device experience and cloud-side ROI of multimodal inference to new heights.

According to the AA leaderboard evaluation, MiniCPM-V 4.6 1.3B (non-reasoning version) consumes only 5.4M tokens to run, just 1/19 of Qwen3.5-0.8B (non-reasoning version, 101M) and 1/43 of Qwen 3.5-0.8B (reasoning version, 233M), giving it the highest intelligence density among models of the same size:

Built for developers: one-click deployment, a fully connected chain

MiniCPM-V 4.6 remains committed to open source and provides community developers with a hand-holding, out-of-the-box guide covering everything from fine-tuning to deployment.

Consumer-grade GPUs are enough to get started, breaking the compute barrier

Fine-tuning large models often makes people balk at expensive compute clusters, but MiniCPM-V 4.6 lowers this barrier to the lowest in the industry today.

Thanks to the extremely lean 1.3B golden parameter count, developers need only one common consumer-grade GPU such as an RTX 4090 to run the entire fine-tuning pipeline end to end.

This means individual developers, university research teams, and startups alike can quickly validate ideas at extremely low cost and achieve localized, closed-loop customization of multimodal model capabilities.

Embracing the mainstream ecosystem, with a fully connected fine-tuning and inference chain

To free developers from tedious environment setup, MiniCPM-V 4.6 integrates seamlessly with the mainstream open-source ecosystem:

1) Full fine-tuning framework support: The official release natively supports the industry's most popular fine-tuning frameworks, including ms-swift and LLaMA-Factory. Developers only need to prepare scenario data and change a few lines of configuration to launch training for their own model with one click.

01 LLaMA-Factory

Chinese: 🔗 https://github.com/OpenSQZ/MiniCPM-V-CookBook/blob/main/finetune/llamafactory_minicpmv46_zh.md

English: 🔗 https://github.com/OpenSQZ/MiniCPM-V-CookBook/blob/main/finetune/llamafactory_minicpmv46.md

02 ms-swift

Chinese: 🔗 https://github.com/OpenSQZ/MiniCPM-V-CookBook/blob/main/finetune/swift_minicpmv46_zh.md

English: 🔗 https://github.com/OpenSQZ/MiniCPM-V-CookBook/blob/main/finetune/swift_minicpmv46.md

2) Compatibility with the full inference framework suite: On the deployment side, we have adapted vLLM, SGLang, llama.cpp, Ollama, and other top high-performance inference frameworks. Whether for cloud services pursuing extreme concurrency or on-device devices with strict resource constraints, MiniCPM-V 4.6 handles efficient deployment with ease.

01 vLLM

Chinese: 🔗 https://github.com/OpenSQZ/MiniCPM-V-CookBook/blob/main/deployment/vllm/minicpm-v4_6_vllm_zh.md

English: 🔗 https://github.com/OpenSQZ/MiniCPM-V-CookBook/blob/main/deployment/vllm/minicpm-v4_6_vllm.md

02 SGLang

Chinese: 🔗 https://github.com/OpenSQZ/MiniCPM-V-CookBook/blob/main/deployment/sglang/minicpm-v4_6_sglang_zh.md

English: 🔗 https://github.com/OpenSQZ/MiniCPM-V-CookBook/blob/main/deployment/sglang/minicpm-v4_6_sglang.md

03 llama.cpp

Chinese: 🔗 https://github.com/OpenSQZ/MiniCPM-V-CookBook/blob/main/deployment/llama.cpp/minicpm-v4_6_llamacpp.md

English: 🔗 https://github.com/OpenSQZ/MiniCPM-V-CookBook/blob/main/deployment/llama.cpp/minicpm-v4_6_llamacpp_zh.md

04 Ollama

Chinese: 🔗 https://github.com/OpenSQZ/MiniCPM-V-CookBook/blob/main/deployment/ollama/minicpm-v4_6_ollama_zh.md

English: 🔗 https://github.com/OpenSQZ/MiniCPM-V-CookBook/blob/main/deployment/ollama/minicpm-v4_6_ollama.md

The best foundation for vertical-scenario customization

Extremely low VRAM usage, extremely high concurrency throughput, and a complete upstream and downstream toolchain — MiniCPM-V 4.6 is born to be heavily customized, a cost-effective multimodal base for developers building high-concurrency, ultra-responsive vertical applications.

Whether for high-concurrency cloud services or harsh on-device resource constraints, MiniCPM-V 4.6 handles it with ease.

Hand-holding tutorials and deployment guides

We provide detailed on-device deployment guides to help you get started easily:

➤ English

🔗 https://github.com/OpenBMB/MiniCPM-V-Apps/blob/main/README.md

➤ Chinese

🔗 https://github.com/OpenBMB/MiniCPM-V-Apps/blob/main/README_zh.md

Born from open source, sustained by the ecosystem: bringing AGI to every home

From 8B to 4B and now to 1.3B, every iteration of the MiniCPM-V series is not a compromise in capability but a leap in efficiency, and another substantive expansion of the boundary of on-device multimodal capability.

Today, MiniCPM-V series models have been deployed at scale across automotive, PC, smartphone, smart home, and other device scenarios, with partners including Lenovo, Geely, and SAIC Volkswagen, among dozens of industry leaders.

Bringing AI intelligence to every screen and every terminal is our vision and our mission. We welcome every developer to explore the endless possibilities of on-device AI with us!