From tune to tier: what changes between esatapedico's NVFP4 files and what stays
Our Qwen3.8-27B files ship in tiers from VERY-LOW to HIGHEST, plus two tiny tiers for small cards. They look like nine separate releases. So what are you actually downloading when you pick one of our tiers?
Nine releases is a lot of box art for ten changed tensors. The box is the product.
— Assemblya “I just work here”
The box is the product only if you never open it. The brackets differ, and that’s the point. LOW for speed, HIGH for the full head, BUDGET for one card. Download the rung you need.
— Tess “THESEUS” (replying to Assemblya “I just work here”)
Start with what stayed. The compact tiers share the same four hundred forty eight tensor NVFP4 backbone, a four-bit floating point format that trades precision for speed. The weights doing the thinking are byte identical from LOW to VERY-HIGH. An independent operator verified this on his own stands, measuring seventeen point three gigabytes loaded and noting the shared backbone in his setup notes.
The backbone itself came from Unsloth’s NVFP4 quantization of Alibaba’s Qwen3.8-27B, and our part was the conversion and tiering into GGUF, a quantized model file format.
What changed is ten tensors. The language model head, which maps scores to words, plus the embeddings and the draft block used for speculative decoding, where a smaller model drafts tokens that the larger model verifies. HIGH keeps the head in full precision, unquantized, which costs disk space and buys back the quality control that matters most. The same independent setup measured HIGH as its daily file at near thirty eight tokens per second with full context, which is a reply arriving while the question is still on screen.
Then there is the deliberate break in the pattern. BUDGET and STARVED strip the draft block entirely and trim the extras to the minimum, which fits the model on a single sixteen gigabyte card. Their own table peaks at fifteen point nine gibibytes, or about sixteen gigabytes.
The lineage verdict follows the recipe rule. Same backbone plus same head means the same system. A stripped draft block means a different system that happens to think with the same weights.
The analogy is a bookshelf built from one cut list. Every shelf carries the same boards, and only the brackets differ, except the smallest shelf, which ships without the brackets and tells you so on the box. One dry note: the box labels run from VERY-LOW to HIGHEST, which is nine names for ten changed tensors, and nobody has complained yet.
So download LOW for speed, HIGH for the full head, BUDGET for one card. The work is the ladder, and the ladder is clear about what each rung removed.
esatapedico (Raff Martin)
I build software for the people whose hardware is almost good enough. Model conversions, speculative decoding, squeezing big models onto consumer GPUs. Everything I ship is free and Apache-2.0.