v2.5: runs all 5 Pingala-8 models, including Qwen3.5

The one-click build now runs all 5 Pingala-8 models: Qwen3-4B, Llama-3.1-8B, Qwen3.5-4B, Qwen3.5-2B and Qwen3.5-0.8B.

What changed

• New runtime: llama.cpp 7ac59a6 + the released CUDA patch (patches/llama_cuda_vs_7ac59a6.patch) + patches/q35_loader_count_delta.patch. The loader now checks that the number of coded tensors equals the file's vr.n_linear instead of assuming 7 per layer. The hybrid Qwen3.5-2B and Qwen3.5-0.8B models have 6 in each DeltaNet layer, so earlier one-click builds cannot load them.
• New launchers, with the same safe download as v2.4.1 (download to <name>.part, check bytes and sha256 from Hugging Face, then rename; 3 tries):
  • Windows: run_qwen35_4b.bat, run_qwen35_2b.bat, run_qwen35_0_8b.bat
  • Linux: ./run.sh qwen35-4b, ./run.sh qwen35-2b, ./run.sh qwen35-0.8b (or run_qwen35_4b.sh, run_qwen35_2b.sh, run_qwen35_0_8b.sh)
• README.txt lists the 5 models and their download sizes.

What did not change

• Same build settings as v2.4.1: Windows CUDA 12.8 (MSVC, GitHub Actions, architectures 86;89;120); Linux CUDA 13.2 (architectures 75-virtual;80;86;89;120). Same CUDA libraries, env.bat settings and llama-server command line.
• The models are unchanged. Model files already in models are kept.
• The loader change keeps the earlier files working: on the rebuilt runtime Qwen3-4B gives the same WikiText-2 perplexity ([73]16.6413) and token generation at 1.01x of the released build.

Tested (RTX 5070 Ti Laptop GPU): on Linux and on Windows (cmd.exe), each of the 5 models loads in this build and answers a 30-token test prompt. A download cut off mid-file was retried and passed the sha256 check on both systems.

Model	Download	Windows	Linux
Qwen3-4B	1.24 GB	run_qwen3_4b.bat	./run.sh qwen
Llama-3.1-8B (base model, Built with Llama)	2.39 GB	run_llama31_8b.bat	./run.sh llama
Qwen3.5-4B (text only)	1.51 GB	run_qwen35_4b.bat	./run.sh qwen35-4b
Qwen3.5-2B (text only)	0.78 GB	run_qwen35_2b.bat	./run.sh qwen35-2b
Qwen3.5-0.8B (text only)	0.35 GB	run_qwen35_0_8b.bat	./run.sh qwen35-0.8b

Windows: download pingala8-windows-cuda12.zip, unzip, double-click a launcher. The first run downloads the model from Hugging Face; then a chat page opens at http://127.0.0.1:8080.

Linux: download pingala8-linux-cuda13.tar.gz, extract, run ./run.sh <model> and open http://127.0.0.1:8080.

Needs an NVIDIA GPU (RTX 30, 40 or 50 series) and a recent driver. Stock llama.cpp, Ollama and LM Studio cannot load these files.

Demo video and poster (from v2.4-oneclick):

• https://github.com/Genrius-AI/pingala8-llm/releases/download/v2.4-oneclick/Pingala8-demo-1080p.mp4
• https://github.com/Genrius-AI/pingala8-llm/releases/download/v2.4-oneclick/Pingala8-demo-poster.jpg

Paper: https://doi.org/10.5281/zenodo.23481757
Models: https://huggingface.co/Genrius-AI/Qwen3-4B-Pingala8-2bit, https://huggingface.co/Genrius-AI/Llama-3.1-8B-Pingala8-2bit (Built with Llama), https://huggingface.co/Genrius-AI/Qwen3.5-4B-Pingala8-2bit, https://huggingface.co/Genrius-AI/Qwen3.5-2B-Pingala8-2bit, https://huggingface.co/Genrius-AI/Qwen3.5-0.8B-Pingala8-2bit

v2.4.1: safer model downloads

Fixes #2: the one-click launchers no longer reuse a model part that only partly downloaded.

What changed

• run_qwen3_4b.bat, run_llama31_8b.bat and run.sh now download each model part to <name>.part and rename it only after the download has finished.
• Each part is checked against its byte count and sha256 as published on Hugging Face before it is used.
• A part on disk with the wrong size (for example after a closed window, Ctrl+C or a network drop) is deleted and downloaded again, with a message that says so. A failed check is retried up to 3 times.
• Same commands as before. Model URLs, env.bat and the llama-server command line are unchanged.

What did not change

• The models are unchanged: same files, sizes and sha256 on Hugging Face. If you already have complete model files in models, they are kept and not downloaded again.
• The binaries and libraries in both archives are byte-identical to v2.4-oneclick; only the launcher scripts were replaced.

Windows: download pingala8-windows-cuda12.zip, unzip, double-click run_qwen3_4b.bat (or run_llama31_8b.bat). The first run downloads the model from Hugging Face; then a chat page opens at http://127.0.0.1:8080.

Linux: download pingala8-linux-cuda13.tar.gz, extract, run ./run.sh qwen (or ./run.sh llama), open http://127.0.0.1:8080.

Needs an NVIDIA GPU (RTX 30, 40 or 50 series) and a recent driver.

Demo video and poster (from v2.4-oneclick):

• https://github.com/Genrius-AI/pingala8-llm/releases/download/v2.4-oneclick/Pingala8-demo-1080p.mp4
• https://github.com/Genrius-AI/pingala8-llm/releases/download/v2.4-oneclick/Pingala8-demo-poster.jpg

Paper: https://doi.org/10.5281/zenodo.23481757
Models: https://huggingface.co/Genrius-AI/Qwen3-4B-Pingala8-2bit and https://huggingface.co/Genrius-AI/Llama-3.1-8B-Pingala8-2bit (Built with Llama)

Pingala-8 one-click runtime (Windows and Linux)

Ready-to-run Pingala-8 runtime: patched llama.cpp (commit 7ac59a6) that runs the B4 2-bit models. Needs an NVIDIA GPU (RTX 30, 40 or 50 series) and a recent driver.

Windows: download pingala8-windows-cuda12.zip, unzip, double-click run_qwen3_4b.bat (or run_llama31_8b.bat). The first run downloads the model from Hugging Face; then a chat page opens at http://127.0.0.1:8080.

Linux: download pingala8-linux-cuda13.tar.gz, extract, run ./run.sh qwen (or ./run.sh llama), open http://127.0.0.1:8080.

Same file size as llama.cpp IQ2_XXS. WikiText-2 perplexity on Qwen3-4B: 16.64 (IQ2_XXS 28.61, our best IQ2_KT 18.15). On Llama-3.1-8B: 8.44 (IQ2_KT 9.36). It loses to IQ2_KT on Qwen3-14B (9.57 against 9.40). Prompt processing is slower than IQ2_XXS (0.74x).

Paper: https://doi.org/10.5281/zenodo.23481757
Models: https://huggingface.co/Genrius-AI/Qwen3-4B-Pingala8-2bit and https://huggingface.co/Genrius-AI/Llama-3.1-8B-Pingala8-2bit (Built with Llama)