Development ·
MiniCPM releases a 35B multimodal variant with selective expert loading
The public 35B-A3B release documents a runtime strategy that predicts and loads selected experts.
Why it matters
Storage and scheduling may change which multimodal models fit a device’s working memory, expanding the practical local-inference design space.
What to watch
Look for independent device measurements, peak-memory reports and output-quality comparisons using matched settings.
Still uncertain. The documented speed and memory figures are configuration-specific vendor measurements. Complete weights still require storage.
Impact assessment reviewed 2026-10-19.
- Who it helps
- Developers evaluating multimodal inference on constrained hardware.
- Availability
- The October 12 model card offers public weights under Apache 2.0. This is the 35B-total, 3B-active variant; the separately listed 1B model is a different checkpoint.
How it works
Lookahead Expert Scheduling predicts which experts the next layer will need, with fallback loading for misses. OpenBMB reports phone performance under a specific configuration. A small resident-memory figure does not describe the full downloaded model or storage requirement.
MiniCPM-V 4.7 35B-A3B model card ↗MiniCPM-V release and deployment documentation ↗
Compare the complete device configuration
Before comparing this model with another local option, record total download size, peak memory, device, quantization and the same image question. Measure first-response delay as well as sustained output speed.
MiniCPM-V 4.7 35B-A3B model card ↗MiniCPM-V release and deployment documentation ↗
What you can do
- Read the linked original and compare its documented scope with the task you need to complete. Open reference ↗
Where the evidence stops
Hardware speed, storage use and peak memory were not measured here. Vendor results depend on quantization, device and runtime settings.
Sources read. Sources read October 19, 2026. No installation, inference or complete external workflow was executed.
Sources & further reading
Hugging Face · implementation
MiniCPM-V 4.7 35B-A3B model card ↗
Original source selected for its explanation of the subject and access requirements.
OpenBMB · page checked 2026-10-19 · Published 2026-10-12
GitHub · implementation
MiniCPM-V release and deployment documentation ↗
Read the October 12 release and October 15 deployment update; hardware performance remains vendor-reported.
OpenBMB · page checked 2026-10-19
Coverage note: This bounded review does not establish complete discussion coverage. Broad X search required sign-in. Product execution and independent replication remain unverified.
The record over time (1)
Dated source checks and published explanations. An observation date can differ from the original release date.
- · Added to Current
2026-10-12: The public 35B-A3B release documents a runtime strategy that predicts and loads selected experts.
Assessment: Storage and scheduling may change which multimodal models fit a device’s working memory, expanding the practical local-inference design space.