The gguf quantization rules just got replaced by measurements. bartowski's per-tensor maps, out…
the gguf quantization rules just got replaced by measurements. bartowski's per-tensor maps, out september 10, assign precision inside gguf files from measured sensitivity instead of general heuristics.
the 4-bit rule died quietly.
Context
Bartowski's post on Hugging Face of September 10, 2026 says upstream llama.cpp quantization uses model-agnostic heuristics, for example bumping the first and last eighth of layers, and describes a framework that measured per-tensor sensitivity to build per-tensor layout maps for GGUF quantization. The metric is KL divergence against the bf16 model using llama-perplexity on wikitext-2-raw at context 512, lower being better. The work was mainly on Qwen models with validation on others. Findings include that token embeddings dominate sensitivity, depth is U-shaped, and per bit the small attention projections and ffn_up and ssm_out are most sensitive while ffn_gate rarely merits extra bits.
The results are tied to that metric, dataset, context length and the tested models, and are one author's measurements. Replacement within his own maps fits the framing, but the post proposes per-tensor layout maps as an approach, and no evidence was inspected that upstream defaults changed or that upstream adopted them. The 4-bit rule is not a rule named in the text read. Whether the maps are released as files or tooling was not extracted. The 96 hour and 1,000 plus configuration counts come from a secondary snippet. The gguf quantization rules just got replaced by measurements and the 4-bit rule died quietly are the note author's framing and take.
Watch next
- Upstream llama.cpp adoption and released maps.
Sources
Provenance
The note above is reproduced unedited from the original post, first published on Threads on 23 September 2026 at 03:48 IST. Sources are the papers and datasets the note draws on.
View the original post ↗Embed this note
More notes
The air is now being asked to keep its own ledger
the air is now being asked to keep its own ledger: ecmwf’s aifs compo becomes the first ai model to forecast atmospheric composition globally every three hours, cleanair simulates 365 days of pm2.5 over china in ten seconds, and a unified framework maps six pollutants at one kilometer across the whole country. the air now files its own composition report.
read the note →The current is now being asked to draw its own map
the current is now being asked to draw its own map: china’s langya 2.0 predicts six ocean phenomena including internal waves and mesoscale eddies, a deep net called wenhai resolves eddies globally with air sea flux formulas built in, and scripps infers surface currents from the way temperature patterns deform in satellite images. the ocean now files its own circulation report.
read the note →The soil is now being asked to report its own carbon
the soil is now being asked to report its own carbon: a nix color sensor paired with generative data augmentation predicts soil organic carbon without a lab, random forest drives 74 percent of soil health mapping studies, and sentinel 2 tracks five year carbon change across france and italy from 922 samples. the dirt now files its own carbon account.
read the note →