MiniCPM-V 4.6

MiniCPM-V 4.6 is the edge-focused vision-language model that OpenBMB uploaded to Hugging Face on 11 May 2026, and the number worth starting with is its size: 1,300,428,016 parameters in total, weights included. That covers both halves of the system — a SigLIP2-400M perception encoder and a Qwen3.5-0.8B language model, connected end to end. The interesting claim in the model card is not that a small model works, but that this one outscores its own foundation. OpenBMB reports 13 points on the Artificial Analysis Intelligence Index against 10 for the bare Qwen3.5-0.8B it is built on, and says it reaches that at roughly a nineteenth of the token cost; against the same base model in thinking mode, scored at 11, the gap in token cost is given as forty-three-fold. It is also reported above Ministral 3 3B, a model more than twice its size. These are the vendor's figures. Efficiency is where the engineering went. Following LLaVA-UHD v4, the vision path cuts encoding FLOPs by more than half, and the model supports a mixed 4x/16x visual token compression rate, so a developer can trade accuracy for speed inside the same weights. OpenBMB puts token throughput at about 1.5 times that of the base language model. The architecture is 24 transformer layers at hidden size 1024 over a 248,094-token vocabulary, with a 27-layer, 1152-wide vision tower working on 980-pixel inputs at patch size 14; the configuration file declares a 262,144-token position budget. Deployment is the other half of the pitch. OpenBMB publishes adaptation code for all three mainstream mobile platforms — iOS, Android and HarmonyOS — and demonstrates the model on an iPhone 17 Pro Max, a Redmi K70 and a HUAWEI nova 14 in unedited screen recordings. Quantised builds ship in GGUF, BNB, AWQ and GPTQ; vLLM, SGLang, llama.cpp and Ollama are supported, as are the SWIFT and LLaMA-Factory fine-tuning stacks. A separate Thinking variant is published alongside. On 17 May 2026 the company opened a hosted API with a free public key. The licence is Apache 2.0, with no revenue threshold and no registration — the same terms as the rest of the current MiniCPM line, and the reason the model can be shipped inside a commercial product without asking anyone. The repository records 748,775 downloads.

▸Videos

No videos yet.

▸Reviews