antirez ships 465 GB Q2 DeepSeek V4 Pro GGUF for DS4 engine

A new file dropped on antirez's Hugging Face repo: a 465 GB two-bit quantized build of DeepSeek V4 Pro, tagged as the imatrix-0813 instruct build and released under an MIT license. Two AI researchers in our directory circulated the link within a day of the upload.

The file name is a full specification of the quantization recipe. Routed mixture-of-experts weights get the aggressive treatment: IQ2_XXS on the gate and up projections, Q2_K on the down projections. Attention projections, shared experts, and the output layer stay at Q8_0. Token embeddings and the learned router stay at F16, and layer norms remain F32. The design principle, spelled out in the repo README, is that routed experts each process only fractions of tokens per layer, so heavy quantization there costs less average quality than the same treatment applied to the router or the shared decision-making components.

Under the hood this is a build for antirez's DS4 inference engine, an experimental runtime the same author has been publishing alongside the GGUFs. The smaller V4-Flash variants in the repo target 128 GB Mac machines at Q2 and 256 GB-plus systems at Q4. The V4-Pro build at 465 GB sits in a heavier weight class, aimed at operators who have the RAM to hold it.

The upload page itself carries no benchmark numbers against a full-precision Pro model, no throughput figures for this specific file, and no note on how well IQ2_XXS holds up on Pro's expert count compared to Flash's. Anyone shipping this will want to run their own evals before wiring it into anything real. If the recipe carries over as cleanly as it did for Flash, a whole tier of frontier open-weight MoE models starts looking like something a single well-provisioned box can serve, rather than something you rent by the hour.