8 GB file temporarily withdrawn while it gets more healing (rare-word and loop issues found in live testing). The 10 GB and 12 GB files are unaffected.
Qwen3.8-27B Swift 1.5, compressed by llamAmpere to fit a 12 GB GPU with long context and MTP speculative decoding on.
Files in this repo: qwen3.8-27b-Swift1.5-Uncensored-ATX-AI3.gguf plus lsa-qwen38-27b-d2048.gguf. Keep them all in one folder.
| Card | File | Context | MTP | Peak VRAM (measured) |
|---|---|---|---|---|
| 12 GB | qwen3.8-27b-Swift1.5-Uncensored-ATX-AI3.gguf |
262,144 | On | 11,544 MiB |
| 10 GB | qwen3.8-27b-Swift1.5-Uncensored-ATX-AI2 (2.3 bpw) |
212,992 | On | — |
| 8 GB | withdrawn, re-healing | — | — | — |
Quality: drift from full precision (KLD, lower is better). All rows use the same 16,200 positions (coding, agentic and RAG at ~12K ctx), scored against Swift 1.5 BF16. The Unsloth rows are Unsloth's exact UD per-tensor recipes and imatrix applied to Swift 1.5, so every row shares one teacher.
| File | Size | KLD | Top-1 |
|---|---|---|---|
| ATX 12 GB | 11.24 GB | 0.0270 | 95.98% |
| Unsloth UD-IQ3_XXS recipe | 10.93 GB | 0.0396 | 95.33% |
| Unsloth UD-Q2_K_XL recipe | 9.95 GB | 0.0591 | 94.00% |
| ATX AI2 2.3 bpw (10 GB) | 9.14 GB | 0.0593 | 94.06% |
| Mirai S | — | 0.0745 | — |
The 12 GB file has 54% less drift than UD-Q2_K_XL. The 2.3 bpw file matches UD-Q2_K_XL in a file 19% smaller.
Shorter answers. On LiveCodeBench v6 (the same 100 problems), the 12 GB file used 33% fewer tokens than UD-Q2_K_XL: median 4,608 vs 8,932 per answer, 95% range 22–44% fewer. It was shorter on 74 of 100 problems.
Coding score, one seed (directional, not statistically significant). LiveCodeBench v6, strict grader: ATX 12 GB 89%, UD-Q2_K_XL 87% (run on llamAmpere), Qwen BF16 published 90.3%.
Fit. With standard attention, UD-Q2_K_XL cannot run full context with MTP on a 12 GB card (estimate). The ATX 12 GB file runs the full 262K context with MTP.
Running the small-card files requires llamAmpere 0.7 (coming, not yet released): LSA attention + host-offloaded KV. The --preset commands below will work once 0.7 ships. Download the sidecar files into the same folder as the model:
# 12 GB
./llama-server -m qwen3.8-27b-Swift1.5-Uncensored-ATX-AI3.gguf --preset 12dense
# 8 GB
Sidecars: lsa-qwen38-27b-d2048.gguf (LSA index map, both sizes). For 8 GB only: mtp-dm.gguf, head-ait2.gguf (draft overlays) and head-exact-f16.gguf (exact rescoring head).
SHA-256:
qwen3.8-27b-Swift1.5-Uncensored-ATX-AI3.gguf d28a3c00f53ce95566df6478b9733d570e8143d5cca669ff057465f1e26812f5lsa-qwen38-27b-d2048.gguf 5489a4f2f6e9169c57db7344afdebdca243c3d438d8e70197c52a6c95244aa68mtp-dm.gguf 4a8e9a5cfb6b1395088bd88b26400b504e47a10fad5a109e4cbee3276f9d5caehead-ait2.gguf 673687fe741bf65c34ebdc73c53248de60f82184cd3a0138b320acfe3d811e5dhead-exact-f16.gguf f812690e5925605edafb2f364e1b40ef89756856c7b6e16ba1e4515bc80d1227Qwen/Qwen3.8-27B