面壁智能端侧多模态 MiniCPM-V 4.7 发布:3GB 内存跑出 13+ token/s
Quick Answer
Thinking Machines Lab launched the open-source Inkling model with 975 billion parameters, enabling efficient customization by activating only a subset of parameters.
Quick Take
This challenges the closed-source AI landscape dominated by major players, offering highly customizable open-weight solutions.
Key Points
- Inkling features a mixed expert architecture for enhanced efficiency.
- The model aims to disrupt the closed-source AI market.
- It allows for significant customization without compromising performance.
- Launched by former OpenAI CTO Mira Murati's new venture.
- Total parameter count reaches 975 billion.
📖 Reader Mode
~1 min read面壁智能当天推出新一代端侧多模态大模型 MiniCPM-V 4.7,一口气给出 35B-A3B 与 1B 两种规格。其中 35B-A3B 是 MiniCPM-V 家族里首个 MoE(混合专家)模型,并依托全新的 LES"前瞻专家调度"推理加速技术,成为业界首个能在端侧无损、流畅运行的开源 MoE。

最抓眼球的数字是落地表现:在 iPhone 17 Pro Max 上,它仅占用 3GB 内存,便跑出 13+ token/s 的解码速度。这意味着无需把数据送上云端,旗舰手机本地就能扛住多模态推理,隐私与延迟双重受益。
跨平台通吃,微调方案一并开源
在适配面上,4.7 同时覆盖 iOS、安卓与鸿蒙三大系统,并兼容 Intel、AMD、Arm、瑞芯微、后摩等芯片平台,把移动端与本地计算的场景一并纳入。面壁还同步放出 Self-Replay Fine-Tuning(SRFT)微调方案,帮开发者以更低成本扩展能力、抑制遗忘,方便做二次开发。
性能对比同样不弱:35B-A3B 在 Instruct 模式下,以更短的推理耗时拿下比 Qwen3.6 35B-A3B 更高的综合得分;1B 版本相较上代 4.6 基础能力继续爬升,综合分与推理效率也都明显压过 Qwen3.5 0.8B。在端侧模型越卷越小的当下,面壁这代把"大模型 MoE 化 + 手机本地跑"的门槛又往下压了一截。
— Originally published at news.aibase.com
Want this in your inbox every morning?
Daily brief at your local 8am — bilingual EN/中文, free.
More from WebSearch (Tavily)
See more →全球AI芯片峰会,9月上海见!
The 2026 Global AI Chip Summit will take place in Shanghai on September 22-23, focusing on the evolving AI chip landscape, including the shift from training to inference, the rise of diverse chip technologies, and the restructuring of industry competition. Notable speakers include experts from leading universities and companies, discussing advancements in AI chip architecture and applications.