Alberto.Codes

HomeAboutExperienceProjectsPublicationsBlogContact

Publications

Artifacts I have published, with what I contributed, what they derive from, and the evidence behind them. Every entry names its base model and license, because a derivative work is not the same as an original one.

Quantized model

2026-08-29

gemma-4-31B-it-fit24gib-GGUF

Google's own 4-bit build of Gemma 4 31B already fits a 24 GiB card at 16.44 GiB, so fit was never the claim. This 14.92 GiB pack ties it on four held-out benchmarks and wins one, and the freed bytes buy served context: 86,016 tokens of text against 65,536, and 73,728 against 49,152 with an image aboard, behind a measured vision bound and a converted projector sidecar.

Base model

google/gemma-4-31B-it-qat-q4_0-unquantized

Relation

quantized

License

Apache 2.0 · Gemma 4 license note

My work

The per-layer sensitivity map, the two-arm solve, the pack, the projector conversion, the serve ladders, the vision and real-GUI campaigns, and the evaluation record. The weights and the projector are Google's.

Quantized model

2026-08-22

NVIDIA-Nemotron-3.5-Lightning-30B-A3B-fit16gib-GGUF

A 30-billion-parameter mixture-of-experts model solved to serve entirely on a 16 GiB card at 16k context. llama.cpp's compact quant types are locked out of the expert stacks that hold 93% of the parameters, so the shelf's smallest build is 17.54 GiB; this 15.76 GiB pack beats it on both ruled damage metrics and holds four leads and a tie on the fixed task slice.

Relation

quantized

License

OpenMDW 1.1

My work

The stack-keyed sensitivity maps, the mixed-precision recipe, the pack, the serve validation, and the evaluation record. The weights are NVIDIA's; the importance matrix is bartowski's, linked at a pinned revision with credit.

Quantized model

2026-08-11

Llama-3_3-Nemotron-Super-49B-v1_5-fit24gib-GGUF

A 49-billion-parameter model fitted to a single 24 GiB RTX 4090 by measuring per-layer quantization damage and solving the bit allocation against the budget, instead of applying a preset. At the same file size it drifts less from the full-precision original than the standard community quantization, and ties it on five capability benchmarks.

Relation

quantized

License

NVIDIA Open Model License · Llama 3.3 Community License

My work

The sensitivity map, the mixed-precision recipe, the pack, and the evaluation record. The weights are NVIDIA's. Built with Llama.

© 2026 Alberto Nieto. All rights reserved.