Opt.Gear: On-Device Language Models Trained on 0.5T Tokens
Introducing the open-source Opt.Gear (270M/1B) on-device text generation models, designed from the outset for environments with limited SRAM, bandwidth, and power.
Read article →
Opt.Gear-QAT: Preserving On-Device Performance with QAT
Introducing the Quantization-Aware Training recipe for Opt.Gear-1B, combining Hessian-based scale initialization with LSQ.
Read article →
Opt.Gear-1M: A Generative Language Model Running on an MCU
Introducing Opt.Gear-1M, a language model with approximately 1M parameters deployed directly on an STM32 Cortex-M7, achieving autoregressive generation at 20 tokens/s.
Read article →More from the team
- Generative inference on a $60 board: Gear-1M on the STM32H747Engineering · July 31, 2026
Publishing with each release
- Gear-VLM-1B — vision-language on the 1B backboneAugust 2026
- Gear-1B-MTP — multi-token prediction on edge acceleratorsAugust 2026
- FunctionGear — structured function calling at the edgeAugust 2026