Reachy Mini: the $300 open source robot you can actually hack — Andres Marafioti, Hugging Face

Andres Marafioti, lead of multimodal research at Hugging Face, presents Reachy Mini, a $300 open-source robot designed to democratize robotics hacking and voice AI experimentation. The talk covers both the product vision and a technically detailed account of optimizing a text-to-speech (TTS) pipeline to production-grade latency.

The motivation for Reachy Mini stems from Marafioti's critique of the current robotics landscape: even the most capable humanoid robots remain in the $50,000–$100,000+ range, are closed systems, and are designed to imitate humans rather than invite creative reinvention. Reachy Mini deliberately avoids human-like form to encourage novel interaction paradigms. Two models exist: a $450 version with a Raspberry Pi and battery, and a $300 version without, with the cheaper unit sold in bulk to high schools and universities. The robot ships unassembled so that users gain full understanding of its hardware and can repair any component. Hugging Face has shipped 7,500 units, and talking to the robot is the most-used app by a significant margin.

The core software stack for voice interaction is a speech-to-speech pipeline Marafioti has maintained at Hugging Face for two years. It uses Parakeet for transcription (running every 150 milliseconds with partial results streamed back to the robot mid-sentence), Qwen 3.5 27B as the LLM for response generation, and Qwen3-TTS (referred to as Coqui 3 TTS in context) for synthesis. The infrastructure runs on Hugging Face Inference Endpoints behind a load balancer, with LLM endpoints deliberately separated from conversation nodes because GPU utilization is highly variable depending on how much individual users speak.

The centerpiece of the technical content is Marafioti's two-week effort to bring Qwen3-TTS from 0.8x real time (generating 1 second of audio in 1.2 seconds) to 5.8x real time (generating 1 second of audio in roughly 170–200 milliseconds). Three root causes were identified and fixed. First, the model lacked streaming, requiring the full audio chunk to be generated before delivery; adding streaming allowed audio to begin flowing before synthesis completed. Second, the autoregressive model performed 500 decoding steps per audio packet, each requiring a CPU–GPU round trip to coordinate state; this is a known bottleneck in PyTorch eager mode. Third, the model used a dynamic KV cache that resized based on input length, which blocked CUDA graph capture. Replacing the dynamic KV cache with a static KV cache (accepting higher up-front RAM usage) enabled CUDA graph captures, eliminating CPU–GPU synchronization overhead and allowing the entire forward pass to execute within the GPU. The combined result: time-to-first-audio dropped from several seconds to under 200 milliseconds, and throughput reached 5.8x real time. The optimized implementation is open-source under the name "faster-Coqui-3-TTS" and has seen daily usage since release approximately two months before the talk.

Marafioti emphasizes that even after achieving sub-200ms model latency, infrastructure round-trip times (network, load balancer, echo cancellation, tool dispatch) roughly equal the model latency in the full system, underscoring that voice agent builders must account for end-to-end latency, not just model benchmarks.

The talk closes with a demonstration of vibe-coding: Marafioti one-shot prompted Claude Code with the Reachy Mini repository to generate the live demo running during the session, requiring no manual coding. He frames the project as a bet on communal development of human-robot interaction patterns before any single company can lock them down.