Compaxis SQ4 is a re-engineered 4-bit GGUF quantization that closes 21–78% of the quality gap back to FP16 compared with the standard Q4_K_M everyone else ships. Drop straight into llama.cpp. Runs on the consumer GPU you already have.
Compaxis SQ4 uses a more efficient compression algorithm than the standard Q4_K_M every other GGUF ships with. The result is a 4-bit file that's measurably closer to FP16 — at the same size, same speed, and same loader.
Across 7 models the average Kullback–Leibler divergence vs FP16 drops by 52% — and by up to 78% on Gemma 4 E4B. Same file format, same loader, just better numbers.
Decode throughput is within a few percent of plain Q4 on every model. You don't pay for the quality with tokens-per-second.
Compaxis ships standard GGUF files. They load in unmodified llama.cpp, Ollama, LM Studio, or anything else that already reads GGUF. No new runtime to install.
KLD-WT is the Kullback-Leibler divergence vs FP16 measured on WikiText. Lower is better. The right column is how much of the Q4_K_M gap to FP16 Compaxis SQ4 closes — across the production model line.
| Model | Q4_K_M KLD | Compaxis SQ4 KLD | Gap closed vs FP16 |
|---|---|---|---|
| Qwen 2.5 0.5B Instruct | 0.034 | 0.027 | +21% |
| Qwen 3 1.7B | 0.121 | 0.050 | +59% |
| Gemma 4 E4B | 0.264 | 0.057 | +78% |
| Qwen 3 8B | 0.053 | 0.031 | +42% |
| Qwen 3 14B | 0.058 | 0.028 | +51% |
| Gemma 4 26B-A4B (MoE) | 3.267 | 0.992 | +70% |
| Qwen 3.6 27B | 0.034 | 0.020 | +41% |
Methodology: each model's FP16 GGUF is the reference. Q4_K_M and Compaxis SQ4 are built from the same FP16 source, then KL divergence is measured against FP16 token-by-token across a fixed WikiText sample.
Install the Hub, pick a model, run. Compaxis SQ4 GGUFs work anywhere a standard GGUF works.
One signed installer for Windows 10 / 11 (x64). Ships free with the Qwen 3 1.7B SQ4 model so you can prove it on your hardware before you pay. Linux support follows in a later release.
Upgrade to Pro to unlock SQ4 of the larger production models — Qwen 3 8B / 14B / 27B, Gemma 4 E4B and 26B MoE. Pro is one subscription, every Pro model included.
The GGUF file loads directly in llama.cpp, Ollama, LM Studio, KoboldCpp, or anything else that reads GGUF. No new runtime, no API to integrate.
Download the Hub. Run the free Qwen 3 1.7B model. Compare it to the same model at Q4_K_M side-by-side. Upgrade when you want the larger models — or don't.