Downloads · 30 days
104
14% of all-time downloads
Qwen3.8-Flash-Next-Simple-Q3_K
IHaveNoClueAndIMustPost/Qwen3.8-Flash-Next-Simple-Q3_K
Qwen3.8-Flash-Next-Simple-Q3_K is a machine learning model from IHaveNoClueAndIMustPost. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
A simpler quantization mix trying to balance speed, size and performance on older hardware. Full Q8 n-gram pipeline.
All-time downloads
729
Public
Repo size
122 GB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.gguf122 GB · 100%
At a glance
- Access
- Public
- Created
- Sep 4, 2026
- Updated
- Sep 14, 2026
- SHA
- 7ee3c667
Base models
Model card
From the Hugging Face model README
A simpler quantization mix trying to balance speed, size and performance on older hardware. Full Q8 n-gram pipeline.
Size<br> 114,245 MiB · 5.42 bpw · 6 shards (core 60.9 GiB + 50.7 GiB table on disk). <br> Approx 65-70GiB total of RAM and/or VRAM needed for 131K context and Q8/F16 cache with n-gram offloading enabled. <br> MTP not included, use whichever standalone version fits best. Though currently <code>--ngram-mod</code> drafting and variants seems to work better in most cases.<br> <b>Addendum</b>: Added a standalone MTP GGUF in simple Q4_0 format. With F16 k/v cache it seems to give a noticeable, but not overwhelming, boost. However prompt processing takes a bit of a hit so depending if your weakest link is PP or TG it might be better to just stick to basic n-gram draft.
Recipe<br> (base Q8_0): signal path → Q8_0 · gate/up → Q3_K (edge blk 0–3 / 40–47 → IQ4_NL) · down → IQ4_NL · n-gram table → Q8_0.
PPL<br> 4.6587 ± 0.028 (WikiText-2, 512 ctx) = +1.49% vs the 4.5903 full BF16.
- Created
- Sep 4, 2026
- Updated
- Sep 14, 2026