A 35B MoE Vision AI Learns to Fit a Phone While a $7 ESP32 Hijacks an iPhone

A 35B MoE Vision Model Squeezes Onto a Phone

On October 12, 2026, OpenBMB released MiniCPM-V 4.7, an Apache-2.0 multimodal family whose 35B-A3B variant is the series' first mixture-of-experts model. Its Lookahead Expert Scheduling prefetches experts across layers, which combined with mixed-precision quantization and DSpark speculative sampling, lifts SSD-offloaded decoding on a 16GB iPad Pro M4 from 2.97 to 17.39 tokens per second, a 5.86x jump OpenBMB reports in its own testing.

The headline number — 13+ tokens per second using 3GB of memory on an iPhone 17 Pro Max — is self-reported by OpenBMB and has not been independently reproduced yet, per explainx.ai's writeup. Day-0 support covers Rockchip, Intel, AMD and Arm chips plus llama.cpp, vLLM and Ollama, so this isn't locked to Apple hardware.

If the numbers hold up outside the vendor's lab, it's a meaningful step: MoE models have mostly stayed server-side because routing overhead eats the latency gains edge devices need.

Google's EmbeddingGemma 2 Puts Multimodal Search in 191MB

Google DeepMind launched EmbeddingGemma 2 this week, an open-weight embedding model that maps text, images, video frames and audio into a single vector space. At 740M parameters, Google says the text-only encoder needs roughly 191MB of active RAM and the full multimodal stack about 567MB on a Pixel 11 Pro — a vendor-reported figure, not independently benchmarked.

The model does zero-shot intent routing without fine-tuning, matching user input directly against labels in milliseconds. Google paired the release with Google AI Edge Foresight for Mac, an offline meeting-notes app built on the same local-first stack.

For developers, the pitch is fewer chained models: one small encoder instead of separate captioning, speech-to-text and text-embedding stages, each of which eats latency and battery on a phone.

A Kids' Voice Assistant Runs Entirely on an ESP32-S3

Developer Anankush Pandit released Ivy AI, an offline voice assistant aimed at children that runs its full pipeline on an ESP32-S3 with 8MB PSRAM and 16MB flash, as covered by Adafruit on October 12, 2026. The WakeNet9 wake-word model listens for 'Hey Ivy,' then a Conformer-based CTC engine transcribes speech before a 19.7-million-parameter language model, loaded into just 6.5MB, generates a reply.

Per a more detailed report, the model produces 5.4 to 5.9 tokens per second, meaning a full spoken exchange takes 14 to 17 seconds end to end — noticeably slower than a phone, but with zero network round trip.

That tradeoff is the whole point of the project: no audio ever leaves the device, which matters more for a kids' toy than shaving seconds off response time.

A $7 ESP32 Lets an AI Agent Drive a Real iPhone

Developer onehumanbeing published esp32-iphone-agent on GitHub on October 11, 2026, and the repo picked up over 1,100 stars within a day. The board uses AirPlay to mirror the iPhone's screen as the agent's 'eyes' and acts as a USB HID device to send taps and swipes as its 'hands,' with any OpenAI-compatible model deciding what to tap.

Worth being clear about the honest caveat here: the model itself runs in the cloud through an API call. The ESP32 is a $7 bridge, not the brain. What makes it a notable maker build is that it lets a cloud LLM physically operate a real phone's touchscreen rather than scripting against an SDK — closer to how a person would use the device.

ESP32 Bit Pirate Gains 5GHz Wi-Fi and 802.15.4 Radios

Geo (@geo-tp) shipped version 0.3 of ESP32-Bus-Expander on October 12, 2026, adding 5GHz Wi-Fi and 802.15.4 radio support to the ESP32 Bit Pirate testing tool, per Adafruit's writeup. Previous versions were limited to 2.4GHz, so this is a real expansion of what the board can sniff and test.

It's not an AI release, but it's the same ESP32 family that's quietly become the default chip for the on-device voice and vision projects running elsewhere in this dispatch — and tooling like this is part of why that ecosystem keeps growing.

A busy day for the small end of the spectrum: one lab-benchmarked MoE model claiming phone-grade speed, one wake-word microcontroller doing the whole voice loop with no cloud at all, and a $7 board making a cloud agent's commands land as real taps on real glass.