An $8 ESP32-S3 now runs a 28.9M-parameter LLM fully offline
nikola_mr64990 · x · 2026-07-29
- A developer fit a 28.9M-parameter LLM onto a regular ESP32-S3 for about $8, running fully offline and drawing roughly the power of a single LED.
- The chip’s constraint was memory: 512 KB fast SRAM and 16 MB flash. The key trick was moving the large embedding table into flash and memory-mapping it.
- Only about 450 bytes per token stay active, so the working set fits in fast SRAM.
- The demo already powers a minimalist voice-controlled remote over Bluetooth LE, showing how far tiny local AI hardware can go.
More from Embodied
- DGX Spark arrives as a local AI workstation for LimestoneHQ experiments — alex_verem · 2026-07-29
- Embodied AI Startup Dream Machines Announces Angel Round Funding — DominiqueCAPaul · 2026-07-29
- US Airlines Ban Humanoid Robots from Flights Citing Battery and Safety Risks — carlosdponx · 2026-07-29
- Musk Claims FSD is Part of the Singularity in Recent Retweet — elonmusk · 2026-07-29
- Codex writes camera firmware and streams a 40-gram stereo rig over Wi‑Fi — yacineMTB · 2026-07-29
- Neuralink says trial participants can steer a powered wheelchair with thoughts — elonmusk · 2026-07-29