Why One Bug Fix Made My Model the Most Downloaded Japanese LLM in MLX Format

I publish MLX-converted Japanese language models on Hugging Face under the handle tocchitocchi. This week, cumulative downloads across my 17 published models crossed 9,000. I wasn't expecting that number. More importantly, I wasn't expecting who was downloading what.

The breakdown nobody predicted

Three model families make up my portfolio:

• Nemotron series (NVIDIA's Japanese-enhanced models): 4,205 downloads — 46% of all downloads
• GPT-OSS-Swallow series: 2,666 downloads — 29%
• Qwen3-Swallow series: 2,361 downloads — 25%

Nemotron dominates by a wide margin. My first assumption was model quality — Nemotron is strong. But quality alone doesn't explain a 2x lead over similarly capable models.

The hypothesis: it wasn't the model, it was the README

During MLX 4-bit conversion of the Nemotron series, I ran into a compatibility issue that produced incorrect output on Apple Silicon. After working through it, I documented the problem and the fix directly in the model card README.

My hypothesis: I may have been one of the few people to publish this specific fix publicly.

When a developer searches for "Nemotron MLX" or hits the same issue, my model card surfaces in the results. The README answers the question they're actually asking: does this work on my machine, and if not, why? That trust converts to downloads.

This is a pattern I've observed across open source: the model with the most useful documentation wins — not the largest, not the newest.

The 120B momentum signal

GPT-OSS-120B shows the highest download momentum in my portfolio at 32.8% of its total coming in recent weeks. The reason is probably hardware scarcity: very few consumer machines can run a 120B parameter model locally. MBP M4 Max 128GB is one of them. Limited working implementations create their own demand.

The machine itself becomes part of the moat.

Three takeaways from 9,000 downloads

1. Document your conversion failures, not just your successes. The gap between "this model works" and "this model works, and here's why it almost didn't" is where trust is built. Most model cards skip the friction entirely.

2. Scarcity compounds. If your conversion requires rare hardware, you're not competing with everyone — you're the only option for a specific subset of users. That subset rewards you with loyalty.

3. Watch momentum, not just totals. Cumulative downloads is a lagging indicator. The models accelerating fastest tell you what the community actually needs right now.

What's next

Behind these model releases is a larger project: I'm building a local-first RAG governance platform called DataTap — designed for enterprise environments where data cannot leave the building. Every benchmark and conversion decision here feeds into understanding which local models are viable for regulated industry deployments.

More on that soon.

All 17 models: huggingface.co/tocchitocchi

Tags: MLX, Japanese NLP, Apple Silicon, Local LLM, Nemotron, Open Source