Why we pay up to $10,000 for improving local models

Our goal is a fully local speech processing pipeline: recognition, diarization, and translation directly on the user's device, without servers. The main obstacle on this path is the quality of local models on consumer-grade hardware. That is why we are launching a bounty program and inviting engineers from all over the world.

How It Works

If you take a local model—whether for speech recognition, speaker diarization, or translation—and make it better than any available open-source alternative, we are ready to reward you. The payout amount depends on how much you have improved the result:

• $100–$1 000 — a noticeable improvement: speedup, reduced memory footprint, or accuracy gains in our scenarios.
• $1 000–$5 000 — a major leap: the model significantly outperforms the best open-source alternative in quality or runs on hardware where it previously couldn't.
• $5 000–$10 000 — a breakthrough: you have essentially rewritten the approach and achieved results multiple times better than any open-source solution. In essence, this is top-tier software that we are partially acquiring.

What We Evaluate

The main criterion is a reproducible result on real-world tasks: speech recognition in calls, speaker diarization, and real-time translation. We look at quality, speed, and resource consumption on typical consumer hardware, rather than benchmarks in a vacuum.

Why We Do This

A small team empowered by AI can ship a product at the speed of a large studio—but breakthrough improvements in models are born from specific individuals with specific ideas. It is more efficient for us to pay fairly for what actually drives the product forward than to spend years trying to build it in-house.

If you are interested, reach out to us via our careers page. Tell us what you have improved and how we can verify it.