MoEspresso Shrinks DeepSeek V4 Flash Coder 32% With No Benchmark

MoEspresso packaged DeepSeek-V4-Flash-0731-Coder at 56.8GB, a 32.63% cut from the 84.4GB build, per a Show HN post. That's the gap between a 204B-parameter coding model fitting on a high-end Mac and one that doesn't. Nobody has checked what the cut costs in output quality.

The tool gets there by deleting about 80B parameters and mixing low-bit quantization schemes: IQ1_S_R4, IQ2_KS and IQ2_K, with heavier q6_K on denser weights. The model keeps 204B parameters total. It only runs under MoEspresso 2.1 or later, and only on Apple MLX.

The demo shown is the pruned model writing a small C compiler locally. The Show HN post is the only source, and it carries no benchmark comparing the pruned build against the original 84GB weights on coding tasks.

A 32% parameter cut with zero quality benchmark isn't a result, it's a claim waiting on a test. Watch for someone to run this against the full model on an actual coding benchmark before trusting the smaller file over the original.