Community conversion of JonathanColetti's Q4_K_M GGUF into Splash's packed affine Q4/group-64 format. The package is approximately 17.37 GB. This is a format conversion, not a new fine-tune or a quality-improved model.
Qwen3.8-27B-Uncensored-Q4_K_M.gguf.4c5e2db039e9325ac7724c8846c71356a24ad1cdfa28002d73ecb6be645f9675.dedf8df68adfb1afeaf7b7480c0a0243108177b4. This standard draft was reused; it was not retrained for the modified target. It is separate from the source GGUF's MTP head.3e6447f082e89cc7f0bc6e5441afd38dfce760ff.All 78 model artifacts were rechecked against their SHA256 hashes before publication. manifest.json records paths, sizes, hashes and upstream revisions.
Tested locally on Apple M5 Pro, 48 GB unified memory, macOS 26.7.1, using Bionic 1.1.10 build 8 / Splash runtime 1.3.0. Initial generation checks were performed on October 31, 2026; API checks were repeated on November 5, 2026.
17 × 23 = 391; √169 = 13.Not established: full 128K input processing, comparative model quality, refusal rates, vision correctness, or end-to-end execution of arbitrary agent tools. A configured context limit is not a full-context benchmark. Upstream evaluation numbers do not describe this requantized package. See test-results.json for the recorded checks.
Download this repository into your LM Studio model directory while preserving manifest.json, target/, draft/, vision/ and tokenizer/. Select the Splash runtime. This packed package does not load as a GGUF or a standard Transformers checkpoint.
For the tested local installation:
lms server start --port 1235 --bind 127.0.0.1
lms load qwen3.8-27b-uncensored-splash \
--identifier qwen3.8-27b-uncensored-splash \
--context-length 131072 --parallel 1 --ttl 600 -y
Model lookup keys can differ after importing the repository; check lms ls and substitute your model key. The explicit API identifier can stay the same.
Client settings:
{
"base_url": "http://127.0.0.1:1235/v1",
"model": "qwen3.8-27b-uncensored-splash",
"api_key": "local"
}
The local tested server does not require an API key; local is a placeholder for clients requiring a non-empty value. TTL 600 unloads the model after ten minutes without use. Actual memory usage grows with context; 128K capacity is not a promise that a full 128K request fits a particular memory limit.
jianchen0311/Qwen3.8-27B-Uncensored-Splash already provides a Splash package based on OrcaRouter's MLX weights. This repository uses a different upstream checkpoint and conversion route. No controlled head-to-head quality comparison has been performed.
Apache 2.0; see LICENSE. Original model: Qwen. Modified GGUF target: JonathanColetti. Splash and DFlash2: Inco AI. Original MLX vision component: mlx-community. Format-conversion tools: publicExcess and contributors. This repository does not imply endorsement by those authors.
vision/model.bin 930 MBtarget/head.bin 715 MBtarget/embedding.bin 715 MBdraft/model.bin 329 MBtarget/layer-0.bin 216 MBtarget/layer-1.bin 216 MBtarget/layer-10.bin 216 MBtarget/layer-12.bin 216 MBtarget/layer-13.bin 216 MBtarget/layer-14.bin 216 MBtarget/layer-16.bin 216 MBtarget/layer-17.bin 216 MBtarget/layer-18.bin 216 MBtarget/layer-2.bin 216 MBtarget/layer-20.bin 216 MBtarget/layer-21.bin 216 MBtarget/layer-22.bin 216 MBtarget/layer-24.bin 216 MBtarget/layer-25.bin 216 MBtarget/layer-26.bin 216 MBtarget/layer-28.bin 216 MBtarget/layer-29.bin 216 MBtarget/layer-30.bin 216 MBtarget/layer-32.bin 216 MBtarget/layer-33.bin 216 MBtarget/layer-34.bin 216 MBtarget/layer-36.bin 216 MBtarget/layer-37.bin 216 MBtarget/layer-38.bin 216 MBtarget/layer-4.bin 216 MBtarget/layer-40.bin 216 MBtarget/layer-41.bin 216 MBtarget/layer-42.bin 216 MBtarget/layer-44.bin 216 MBtarget/layer-45.bin 216 MBtarget/layer-46.bin 216 MBtarget/layer-48.bin 216 MBtarget/layer-49.bin 216 MBtarget/layer-5.bin 216 MBtarget/layer-50.bin 216 MBtarget/layer-52.bin 216 MBtarget/layer-53.bin 216 MBtarget/layer-54.bin 216 MBtarget/layer-56.bin 216 MBtarget/layer-57.bin 216 MBtarget/layer-58.bin 216 MBtarget/layer-6.bin 216 MBtarget/layer-60.bin 216 MBtarget/layer-61.bin 216 MBtarget/layer-62.bin 216 MBtarget/layer-8.bin 216 MBtarget/layer-9.bin 216 MBtarget/layer-11.bin 209 MBtarget/layer-15.bin 209 MBtarget/layer-19.bin 209 MBtarget/layer-23.bin 209 MBtarget/layer-27.bin 209 MBtarget/layer-3.bin 209 MBtarget/layer-31.bin 209 MBtarget/layer-35.bin 209 MBtarget/layer-39.bin 209 MBtarget/layer-43.bin 209 MBtarget/layer-47.bin 209 MBtarget/layer-51.bin 209 MBtarget/layer-55.bin 209 MBtarget/layer-59.bin 209 MBtarget/layer-63.bin 209 MBtarget/layer-7.bin 209 MBdraft/layer-0.bin 187 MBdraft/layer-1.bin 187 MBdraft/layer-2.bin 187 MBdraft/layer-3.bin 187 MBdraft/layer-4.bin 187 MB有文件?提交种子。
提交种子