# Pingala-8 by Genrius AI: full results

> Pingala-8 is an open-source 2-bit weight format for local LLMs in llama.cpp, at the file size of llama.cpp's IQ2_XXS: an 8-state Ungerboeck trellis code over groups of 4 weights, after a Hadamard rotation and LDLQ rounding, run by a CUDA patch to llama.cpp. It compresses existing models without retraining. Files need the patched Pingala-8 runtime (CUDA, NVIDIA RTX 30/40/50); stock llama.cpp, Ollama and LM Studio cannot load them. Apache License 2.0.

Genrius AI. Founded by Harshit Goyal, former Google software engineer and IIIT Delhi alumnus.
Site: https://genrius-ai.github.io/pingala8-llm/
Code: https://github.com/Genrius-AI/pingala8-llm
Models: https://huggingface.co/Genrius-AI
Paper: DOI 10.5281/zenodo.23481757 (https://doi.org/10.5281/zenodo.23481757). The Qwen3.5-2B version 2 and Qwen3.8-27B numbers below are newer than the paper and are not in it.
Short version: https://genrius-ai.github.io/pingala8-llm/llms.txt
Updated: 2026-11-01.

All perplexities are WikiText-2 test, context 4096; lower is better. GB = 10^9 bytes.

## Qwen3.5-2B: six 2-bit GGUF files under 1 GB (November 1, 2026)

72 windows of 4096 tokens, every file on the same windows. "Paired" is Pingala-8 version 2 minus that file, mean nats per token over 70 paired windows (positive means the other file is better), with the number of windows on which version 2 has the lower loss.

| File | Bytes | Size | PPL | Paired against Pingala-8 v2 |
|---|---|---|---|---|
| DASH-Q IQ2_M | 786,259,840 | 0.79 GB | 13.33 | DASH-Q better: v2 wins only 3/70, +0.053 |
| DASH-Q IQ2_XS | 757,096,320 | 0.76 GB | 13.90 | DASH-Q better: v2 wins only 21/70, +0.012 +/- 0.007 |
| Pingala-8 v2 (ours) | 778,189,408 | 0.78 GB | 14.06 | - |
| unsloth UD-IQ2_M | 859,857,152 | 0.86 GB | 14.96 | v2 better on 67/70, -0.062 +/- 0.009 |
| Pingala-8 v1 (ours) | 778,744,192 | 0.78 GB | 15.62 | v2 better on 70/70, -0.104 |
| unsloth UD-IQ2_XXS | 768,270,592 | 0.77 GB | 23.41 | v2 better on 70/70 |

- DASH-Q IQ2_M has the best perplexity of the six. Both DASH-Q files are better than Pingala-8 version 2.
- Model: https://huggingface.co/Genrius-AI/Qwen3.5-2B-Pingala8-2bit (file Qwen3.5-2B-Pingala8-2bit-v2.gguf). Text model only.
- Version 2 needs the patched Pingala-8 runtime (CUDA). The one-click launcher still downloads version 1; for version 2, download the v2 file and pass it with -m.
- Rival files ran on llama.cpp upstream 889f2ec. On our machine (RTX 5070 Ti Laptop GPU) the IQ1_S, IQ2_S and IQ3_S CUDA kernels failed test-backend-ops, so those tensors ran on the CPU. Details: https://genrius-ai.github.io/pingala8-llm/2bit-gguf-nan-perplexity-rtx-50.html
- Page: https://genrius-ai.github.io/pingala8-llm/qwen3-5-2b-under-1gb-2bit.html
- Per-window data: https://huggingface.co/datasets/Genrius-AI/2bit-gguf-wikitext2-eval ; ranking: https://huggingface.co/spaces/Genrius-AI/2bit-gguf-leaderboard

## Qwen3.8-27B: 8.31 GB on one 12 GB GPU (November 1, 2026)

First 20 of 72 windows of 4096 tokens (a time cap on one laptop GPU), every file on the same windows.

| File | Size | PPL | Against Pingala-8 v2 |
|---|---|---|---|
| DASH-Q IQ2_XS | 8.55 GB | 6.54 | DASH-Q better: +0.052 nats/token, on 20/20 windows |
| Pingala-8 v2 (ours) | 8.31 GB | 6.89 | - |
| unsloth UD-IQ2_S | 8.37 GB | 7.34 | ours lower on average, not significant (better on 9/20) |
| Pingala-8 v1 (ours) | 8.31 GB | 7.34 | v2 better on 20/20 |
| IQ2_XXS (llama.cpp, same imatrix) | 8.43 GB | 7.88 | ours lower |

- Speed: 37.7 tokens/s token generation, batch 1, every layer on one RTX 5070 Ti Laptop GPU (12 GB, WSL2, -ngl 99).
- Model: https://huggingface.co/Genrius-AI/Qwen3.8-27B-Pingala8-2bit (8,314,241,600 bytes). Text model only. Needs the patched Pingala-8 runtime (CUDA); not tested with the one-click build.
- Page: https://genrius-ai.github.io/pingala8-llm/qwen3-8-27b-vram-requirements.html

## Results across 7 models (Pingala-8 against llama.cpp IQ2_XXS at the same file size)

| Model | Size | PPL: Pingala-8 vs IQ2_XXS | IQ2_XXS / Pingala-8 | Note |
|---|---|---|---|---|
| Qwen3.5-0.8B | 0.35 GB | 25.97 vs 85.09 | 3.28x lower | |
| Qwen3.5-2B (version 1) | 0.78 GB | 15.62 vs 29.08 | 1.86x lower | version 2 is 14.06, see above |
| Qwen3.5-4B | 1.51 GB | 10.97 vs 14.75 | 1.34x lower | |
| Qwen3-0.6B | 0.28 GB | 41.34 vs 10,791 | 261x lower | IQ2_XXS collapses; IQ2_KT 48.97 in a 3.1% larger file; Pingala-8 generates at 0.70x IQ2_XXS speed |
| Qwen3-1.7B | 0.69 GB | 28.23 vs 47.96 | 1.70x lower | loses to IQ2_KT (20.89, 3.3% larger file) |
| Qwen3-4B | 1.24 GB | 16.64 vs 28.61 | 1.72x lower | |
| Llama-3.1-8B | 2.39 GB | 9.14 vs 12.35 | 1.35x lower | pre-registered; with a BOS at every window 8.44 vs 9.36 for IQ2_KT |

- Qwen3-4B: 8.05 GB to 1.24 GB; at IQ2_XXS size it scores 21% higher on five quizzes (average 50.8 vs 41.9).
- Qwen3-14B: the recipe narrowly loses to IQ2_KT (9.57 vs 9.40).

## Limits

- CUDA only: the patched Pingala-8 llama.cpp runtime on NVIDIA RTX 30/40/50 GPUs. No Mac, AMD or CPU-only path.
- Stock llama.cpp, Ollama and LM Studio cannot load Pingala-8 files. The Qwen3.5, Qwen3-1.7B and Qwen3-0.6B files need runtime v2.5 or later.
- Not the lowest perplexity at 2 bits: DASH-Q is better on Qwen3.5-2B and Qwen3.8-27B; IQ2_KT is better on Qwen3-1.7B and narrowly on Qwen3-14B.
- Prompt processing is 0.74-0.89x the speed of IQ2_XXS. Some IQ2_KT files are smaller (Qwen3.5 rows).
- Qwen3-1.7B and Qwen3-0.6B have no one-click launcher yet. The one-click launcher downloads Qwen3.5-2B version 1, not version 2.

## Models

- https://huggingface.co/Genrius-AI/Qwen3.8-27B-Pingala8-2bit
- https://huggingface.co/Genrius-AI/Qwen3.5-2B-Pingala8-2bit
- https://huggingface.co/Genrius-AI/Qwen3.5-4B-Pingala8-2bit
- https://huggingface.co/Genrius-AI/Qwen3.5-0.8B-Pingala8-2bit
- https://huggingface.co/Genrius-AI/Qwen3-4B-Pingala8-2bit
- https://huggingface.co/Genrius-AI/Qwen3-1.7B-Pingala8-2bit
- https://huggingface.co/Genrius-AI/Qwen3-0.6B-Pingala8-2bit
- https://huggingface.co/Genrius-AI/Llama-3.1-8B-Pingala8-2bit

## Links

- One-click release (Windows and Linux): https://github.com/Genrius-AI/pingala8-llm/releases/latest
- How to reproduce: https://github.com/Genrius-AI/pingala8-llm/blob/main/REPRODUCE.md
- Benchmark dataset: https://huggingface.co/datasets/Genrius-AI/pingala8-benchmarks
- 2-bit GGUF leaderboard: https://huggingface.co/spaces/Genrius-AI/2bit-gguf-leaderboard
- 2-bit GGUF WikiText-2 eval dataset: https://huggingface.co/datasets/Genrius-AI/2bit-gguf-wikitext2-eval
- GPU fit calculator: https://huggingface.co/spaces/Genrius-AI/llm-vram-calculator
- Announcement discussion: https://github.com/Genrius-AI/pingala8-llm/discussions/1
- Paper page: https://genrius-ai.github.io/pingala8-llm/paper.html
- Paper: https://doi.org/10.5281/zenodo.23481757