| --- |
| license: cc-by-4.0 |
| pretty_name: Pingala-8 Benchmarks |
| language: |
| - en |
| tags: |
| - quantization |
| - llama.cpp |
| - gguf |
| - benchmark |
| - llm |
| - model-compression |
| - 2-bit |
| - qwen3 |
| - llama-3.1 |
| - local-llm |
| size_categories: |
| - n<1K |
| configs: |
| - config_name: results |
| default: true |
| data_files: |
| - split: train |
| path: results.csv |
| - config_name: speed |
| data_files: |
| - split: train |
| path: speed.csv |
| - config_name: zero_shot |
| data_files: |
| - split: train |
| path: zero_shot.csv |
| --- |
| |
| # Pingala-8 Benchmarks |
|
|
| These are the measured results behind Pingala-8, our 2-bit weight format for llama.cpp, from Genrius AI. Each number comes from the paper and its logs. We publish the tables so anyone can check, sort and reuse them. |
|
|
| **Headline, with scope:** on Qwen3-4B, at the file size of llama.cpp's IQ2_XXS (about 1.24 GB), Pingala-8 reaches WikiText-2 perplexity 16.64. IQ2_XXS reaches 28.61, and the best IQ2_KT file we built at this size reaches 18.15 (lower is better). The 5-task zero-shot average is 50.83, against 41.86 for IQ2_XXS and 46.52 for that IQ2_KT file. The 8-bit original scores 56.3. No retraining is involved: Pingala-8 compresses the original weights. |
| |
|  |
| |
| The Llama panel of the chart uses llama.cpp's strided mode, where windows after the first start without a BOS token. |
| |
| ## Files |
| |
| | Config | File | What it holds | |
| |:-|:-|:-| |
| | `results` (default) | `results.csv` | WikiText-2 perplexity, file bytes, bits per weight and zero-shot average for Qwen3-4B, Llama-3.1-8B and Qwen3-14B | |
| | `speed` | `speed.csv` | Speed relative to IQ2_XXS on one GPU, with 95% intervals | |
| | `zero_shot` | `zero_shot.csv` | Per-task zero-shot accuracy on Qwen3-4B | |
|
|
| ```python |
| from datasets import load_dataset |
| results = load_dataset("Genrius-AI/pingala8-benchmarks", "results", split="train") |
| speed = load_dataset("Genrius-AI/pingala8-benchmarks", "speed", split="train") |
| ``` |
|
|
| ## Columns |
|
|
| `results.csv` |
| - `model`: the base model. |
| - `format`: the file format, with the runtime in brackets. "Pingala-8 (B4, ours)" is the shipped file, called B4 in the paper. |
| - `file_bytes`: size of the GGUF file in bytes. |
| - `bpw_all`: bits per weight over all tensors, calculated as file_bytes x 8 / the model's parameter count (4,022,468,096 for Qwen3-4B; 4,021,784,576 for Ternary Bonsai 4B). It includes the token embedding and the GGUF metadata. Left empty where we did not calculate it. |
| - `wikitext2_ppl`: WikiText-2 test perplexity (lower is better). |
| - `ratio_to_fp16`: perplexity divided by the fp16 perplexity of Qwen3-4B (12.4567). Qwen3-4B only. |
| - `zero_shot_avg`: mean of the five zero-shot accuracies in `zero_shot.csv`, where measured. |
| - `note`: the evaluation protocol and anything you need to read the row fairly. |
| - `source`: the table in the paper that the number comes from. |
|
|
| `speed.csv`: `tg128_ratio` is batch-1 generation of 128 tokens and `pp512_ratio` is processing a 512-token prompt. Both are given as the mean same-round ratio to IQ2_XXS, with a 95% t-interval. `batched_Nseq_ratio` is decoding N sequences at once (llama-batched-bench, 128-token prompt, 128 tokens per sequence), as a ratio to IQ2_XXS. |
|
|
| ## Protocol |
|
|
| - **Perplexity:** WikiText-2 test (raw, joined), `llama-perplexity` in strided mode, with windows of 4096 tokens (`-c 2049 --ppl-stride 4095`). Qwen3-4B and Qwen3-14B use 73 windows and Llama-3.1-8B uses 70. For Llama-3.1-8B we report two runs. The first uses strided mode, where windows after the first start without a BOS token. The second is a pre-registered rerun with a BOS at every window. The fp16 reference on Qwen3-4B is 12.4567. |
| - **Zero-shot:** `llama-perplexity` task modes (`--hellaswag`, `--winogrande`, `--multiple-choice`), 0-shot, scored by the log-likelihood of each answer's continuation. The sets are HellaSwag validation (1000 items), the WinoGrande debiased file (1267), ARC-Easy and ARC-Challenge validation (570 and 299) and the full MMLU test (14,042). Q8_0 stands in for fp16. Typical standard errors for a 2-bit file: HellaSwag +-3.1, WinoGrande +-1.37, ARC-Easy +-2.06, ARC-Challenge +-2.79, MMLU +-0.40. These continuation scores sit far below 5-shot MMLU numbers, so do not compare them with leaderboards. |
| - **Speed:** `llama-bench -ngl 99 -fa 1 -ctk f16 -ctv f16`, tg128 as `-p 0 -n 128` and pp512 as `-p 512 -n 0`. The runs were one exclusive session on one RTX 5070 Ti Laptop GPU, 10 rounds with rotated order, `-r 5`, and no other GPU process. IQ2_XXS medians were 197.1 tokens/s (tg128) and 7561 tokens/s (pp512). |
| - **Rival files:** we ran GSQ Q2_K and Ternary Bonsai 4B ourselves with the same perplexity command. Bonsai was run on a Q8_0 copy of its 16-bit release. Its listed size is calculated from tensor shapes for its Q2_0 file, which needs the authors' llama.cpp fork. |
| |
| ## Honest limits |
| |
| - **Qwen3-14B is a loss.** IQ2_KT reaches 9.40 against our 9.57 on WikiText-2, and the two tie on C4. We wrote this test down before running it. |
| - **The Qwen3-4B result is post hoc.** B4 is the last of a chain of builds, all evaluated on WikiText-2 during development, so treat its WikiText-2 number as tuned. The C4 text, the Llama-3.1-8B run and the Qwen3-14B run were held out or pre-registered. |
| - **Speed is mixed.** Batch-1 generation runs at 1.096x the speed of IQ2_XXS. Prompt processing runs at 0.74x, and batched decoding of 2 to 8 sequences at 0.18x to 0.34x, with the shipped kernel. All speed numbers come from one laptop GPU. |
| - **Some rivals are better.** Retrained models score better: Ternary Bonsai 4B reaches 11.59 at 1.07 GB, but it is a new trained model, not a compression of Qwen3-4B. Larger files also score better: GSQ Q2_K reaches 14.37 at 1.67 GB, 34% larger than ours. The claim we make is narrower. Among the formats we tested, Pingala-8 is the best way we found to compress the original Qwen3-4B to this size without retraining. |
| - **It needs our runtime.** The files run only on our patched llama.cpp (CUDA). Stock llama.cpp, Ollama and LM Studio cannot load them yet. |
| - **Coverage is small.** We measured three models (4B, 8B and 14B) and one GPU. Zero-shot and speed were measured on Qwen3-4B only. |
|
|
| ## Method in one line |
|
|
| Pingala-8 uses an 8-state trellis code (Ungerboeck) over groups of 4 weights, after a Hadamard rotation and QuIP-style LDLQ rounding. It builds on QTIP, QuIP# and ikawrakow's IQ2_KT. |
| |
| ## Links |
| |
| - Paper: https://doi.org/10.5281/zenodo.23481757 |
| - Code, llama.cpp patch and logs: https://github.com/Genrius-AI/pingala8-llm |
| - One-click Windows build (NVIDIA RTX 30/40/50): https://github.com/Genrius-AI/pingala8-llm/releases/tag/v2.4-oneclick |
| - Models: https://huggingface.co/Genrius-AI/Qwen3-4B-Pingala8-2bit (Apache-2.0) and https://huggingface.co/Genrius-AI/Llama-3.1-8B-Pingala8-2bit (Llama 3.1 Community License, Built with Llama) |
| |
| ## License and citation |
| |
| The tables and chart in this dataset are released under CC-BY-4.0. The models have their own licenses, linked above. |
| |
| ```bibtex |
| @misc{goyal2026pingala8, |
| title = {Small-State Coset Trellis Coding at IQ2\_XXS Size: An Equal-Bytes Study in llama.cpp}, |
| author = {Goyal, Harshit}, |
| year = {2026}, |
| publisher = {Zenodo}, |
| doi = {10.5281/zenodo.23481757}, |
| note = {Genrius AI} |
| } |
| ``` |
| |
