TechnologyTracking
MiniCPM-V 4.7 performance on 3GB devices?
ModelBest's open-source MiniCPM-V 4.7 claims lossless operation on edge devices with 3GB memory, but how do its actual inference speed, energy efficiency, and task accuracy compare with cloud-based large models? Will this breakthrough accelerate the adoption of on-device AI applications (e.g., smartphones, IoT devices) and reshape the industry's deployment landscape?
Watching: 关注第三方独立评测机构对MiniCPM-V 4.7在真实设备上的推理速度、能效和任务精度的测试结果，以及手机厂商、物联网设备厂商的采用公告和行业报告。
Current read
ModelBest released MiniCPM-V 4.7; the 35B-A3B version runs in ~3GB memory with 13-15+ token/s decoding, but evidence lacks direct comparison with cloud LLMs or verified industry impact.
Reports confirm ModelBest released MiniCPM-V 4.7 on October 13, 2026, with 35B-A3B (MoE, 35B total, 3B activated) and 1B versions, running in ~3GB memory with decoding speeds reported as 13+ or 15+ token/s, and claiming superior performance among same-size models. However, all evidence comes from launch coverage; no direct comparison with cloud LLMs on speed, efficiency, or accuracy is provided, nor are adoption announcements or industry-structure changes mentioned. Thus, the reports establish the release and some specs but cannot independently verify the 'lossless' claim or actual impact on edge-AI adoption.
Key developments
- The assessment representation was corrected; evidence and earlier records are retained.
How we read it
Reports confirm the release and specs: the 35B-A3B version runs in ~3GB memory with 13+ or 15+ token/s decoding, but no direct comparison with cloud LLMs is provided.
Reports claim superior performance among same-size models, but lack third-party independent benchmarks or adoption announcements, so edge-AI adoption or industry impact cannot be verified.
Earlier updates
- — Public tracking started; no independent probability estimate yet.; Starting public context: Released Oct 13 in 35B-A3B and 1B versions, using LES expert scheduling, weight layout optimization and visual token compression to run on edge devices with 3GB of memory.
- — ModelBest released MiniCPM-V 4.7; the 35B-A3B version runs in ~3GB memory with 13-15+ token/s decoding, but evidence lacks direct comparison with cloud LLMs or verified industry impact.; Reports confirm ModelBest released MiniCPM-V 4.7 on October 13, 2026, with 35B-A3B (MoE, 35B total, 3B activated) and 1B versions, running in ~3GB memory with decoding speeds reported as 13+ or 15+ token/s, and claiming superior performance among same-size models. However, all evidence comes from launch coverage; no direct comparison with cloud LLMs on speed, efficiency, or accuracy is provided, nor are adoption announcements or industry-structure changes mentioned. Thus, the reports establish the release and some specs but cannot independently verify the 'lossless' claim or actual impact on edge-AI adoption.