ModelBest uploaded a new model, MiniCPM-V-4.7-35B-A3B, to Hugging Face this week, but as of press time has not released a model card (the official technical documentation). The naming itself reveals several key details: "V" stands for Vision, meaning it's a multimodal model supporting images; 35B is the total parameter count; A3B refers to 3B Active Parameters—the hallmark of a classic Mixture of Experts (MoE) architecture, which splits the model into multiple "expert" subnetworks and only calls a subset at inference.

What This Is

ModelBest has long taken a "small but mighty" route—the MiniCPM series previously centered on compact 2B and 4B variants. Jumping straight to 35B total parameters plus MoE is a clear upgrade move. Community feedback suggests that 3B activation means it can run on a single consumer GPU with 24GB of VRAM (such as an RTX 4090) at decent speed—exactly what local-deployment enthusiasts care about most.

Industry View

The open-source community is broadly excited: MoE is turning "running big models locally" from concept into practice. This year, Qwen3-30B-A3B and DeepSeek-V3 are both betting on this track, so ModelBest following suit is moving with the tide.

But there are also cautionary voices. First, no model card means this isn't an official release—it looks more like an early weight leak or a pre-release test build. Quality, licensing, and commercial boundaries all remain unknown. Second, MoE is no longer a blue-ocean track: Qwen3 and DeepSeek hold first-mover advantage, and it won't be easy for ModelBest to differentiate on benchmark leaderboards or in the application ecosystem. Third, "3B activation" sounds lightweight, but 35B total parameters means VRAM usage is still higher than a pure 3B model—the barrier for average users to deploy locally hasn't dropped to "anyone can run it."

Impact on Regular People

  • For enterprise IT: Open-source LLM options keep growing, and so do evaluation costs. The question shifts from "is there a usable model" to "which fits our scenario best," requiring dedicated headcount for benchmarking and compliance review.
  • For individual professionals: Limited impact. Unless you're an AI engineer or experimenting with local deployment, your daily work has no use case for this.
  • For the consumer market: No short-term impact. Consumer users still interact with wrapped products like ChatGPT, Doubao, and Tongyi. Architecture upgrades at the foundation layer need an official release and optimization before they translate into product-level experience.