Qwen3-Coder 30B-A3B Heretic (imatrix GGUF)
This entry is less about a new base model and more about practical deployment: mradermacher/Qwen3-Coder-30B-A3B-Instruct-Heretic-i1-GGUF is a GGUF quant collection (plus the imatrix file) for a coder-focused Qwen3 mixture-of-experts model. If you run models via llama.cpp (or any GGUF-compatible stack), “imatrix” packs are often worth a look because they use an importance matrix to preserve higher-value weights during quantization, which can improve quality at the same size compared to purely static quants.
The model card includes a quant table spanning very small IQ variants up through the more commonly useful options. If you just want a sane starting point on consumer GPUs, the listed Q4_K_M / Q4_K_S variants are typically the best first try, then move up to Q6_K if you have headroom and want extra quality. It also ships an *.imatrix.gguf so you can generate your own quant variants if you have specific performance constraints.
Quick stats from the listing feed: pipeline: text-generation · 1 like · 12,856 downloads.
Source listing: https://huggingface.co/models?sort=modified