Wladastic/Qwen3.8-Flash-Next-Ombro-v13b-GGUF overview
Qwen3.8 Flash Next — Ombro V13b GGUF Mixed precision GGUF quantization of Qwen/Qwen3.8 Flash Next https://huggingface.co/Qwen/Qwen3.8 Flash Next , successor of…
Runs locally from ~1.24 GB disk (4 GB VRAM class GPUs with llama.cpp / guIDE).
Repository Files & Downloads
| File | Type | Quantization | Size | Link |
|---|---|---|---|---|
| Qwen3.8-Flash-Next-Ombro-v13b-00001-of-00009.gguf | GGUF | GGUF | 3.29 GB | Download |
| Qwen3.8-Flash-Next-Ombro-v13b-00002-of-00009.gguf | GGUF | GGUF | 26.82 GB | Download |
| Qwen3.8-Flash-Next-Ombro-v13b-00003-of-00009.gguf | GGUF | GGUF | 12.64 GB | Download |
| Qwen3.8-Flash-Next-Ombro-v13b-00004-of-00009.gguf | GGUF | GGUF | 13.05 GB | Download |
| Qwen3.8-Flash-Next-Ombro-v13b-00005-of-00009.gguf | GGUF | GGUF | 12.94 GB | Download |
| Qwen3.8-Flash-Next-Ombro-v13b-00006-of-00009.gguf | GGUF | GGUF | 13.91 GB | Download |
| Qwen3.8-Flash-Next-Ombro-v13b-00007-of-00009.gguf | GGUF | GGUF | 13.29 GB | Download |
| Qwen3.8-Flash-Next-Ombro-v13b-00008-of-00009.gguf | GGUF | GGUF | 13.59 GB | Download |
| Qwen3.8-Flash-Next-Ombro-v13b-00009-of-00009.gguf | GGUF | GGUF | 1.24 GB | Download |
Model Details
| Model ID | Wladastic/Qwen3.8-Flash-Next-Ombro-v13b-GGUF |
|---|---|
| Author | Wladastic |
| Pipeline | text-generation |
| License | other |
| Base model | Qwen/Qwen3.8-Flash-Next |
| Last modified | 2026-10-15T11:29:10.000Z |
Model README
---
base_model: Qwen/Qwen3.8-Flash-Next
license: other
license_name: qwen-community-1.0
license_link: LICENSE
library_name: llama.cpp
pipeline_tag: text-generation
tags:
- gguf
- qwen
- qwen3.8
- moe
- quantized
- ombro
---
Qwen3.8 Flash Next — Ombro V13b GGUF
Mixed-precision GGUF quantization of
Qwen/Qwen3.8-Flash-Next, successor of
| Quant | Size | Effective BPW |
| --- | ---: | ---: |
| UD-Q4_K_XL | 111.3 GB | 5.03 |
| Ombro V9 | 114.5 GB | 5.18 |
| Ombro V13b | 118.9 GB | 5.38 |
| UD-Q5_K_XL | 158.3 GB | 7.16 |
Compared with UD-Q4_K_XL and Ombro V9
Routed experts
Per layer, gate / up / down. Bold marks a change from V9.
| Layer | UD-Q4_K_XL | Ombro V9 | Ombro V13b |
| --- | --- | --- | --- |
| all other 39 layers | Q4_K / Q4_K / Q5_1 | Q4_K / Q4_K / Q5_1 | Q4_K / Q4_K / Q5_1 |
| 2 | Q5_K / Q5_K / Q8_0 | Q6_K / Q6_K / Q8_0 | Q6_K / Q6_K / Q8_0 |
| 4, 30, 47 | Q4_K / Q4_K / Q8_0 | Q6_K / Q6_K / Q8_0 | Q6_K / Q6_K / Q8_0 |
| 17 | Q4_K / Q4_K / Q5_1 | Q6_K / Q6_K / Q5_1 | Q6_K / Q6_K / Q5_1 |
| 33 | Q4_K / Q4_K / Q5_1 | IQ4_XS / Q6_K / Q5_1 | Q6_K / Q6_K / Q8_0 |
| 35 | Q4_K / Q4_K / Q5_1 | Q6_K / Q6_K / Q8_0 | Q6_K / Q6_K / Q8_0 |
| 38 | Q4_K / Q4_K / Q5_1 | Q4_K / Q6_K / Q5_1 | Q6_K / Q6_K / Q8_0 |
| 46 | Q4_K / Q4_K / Q8_0 | Q4_K / Q5_K / Q8_0 | Q6_K / Q6_K / Q8_0 |
Everything else
| Part | UD-Q4_K_XL | Ombro V9 | Ombro V13b |
| --- | --- | --- | --- |
| Attention and SSM projections | Q8_0 | Q8_0 | BF16 |
| Hyper-connection projections | Q8_0 | Q8_0 | BF16 |
| PLE projections | Q8_0 | Q8_0 | BF16 |
| Norms, router, SSM and inject control tensors | F32 | F32 | F32 |
| Shared experts | Q8_0 | Q8_0 | Q8_0 |
| Token embedding / output | Q8_0 | Q8_0 | Q8_0 |
| Per-layer token embedding (PLE) table | IQ4_NL | IQ4_NL | IQ4_NL |
| Routed experts | 77.0 GB | 80.2 GB | 81.5 GB |
| Total | 111.3 GB | 114.5 GB | 118.9 GB |
Format limits behind these choices
- The
downexperts have 640-wide rows, which K-quants (256-element blocks) cannot
store. Only 32-element block types fit, and between Q5_1 (6.0 bpw) and Q8_0
(8.5 bpw) there is nothing, so down is either Q5_1 or Q8_0.
- The PLE table has 160-wide rows and is stored as
IQ4_NL. It is 28.8 GB, about a
quarter of the model, but each token reads only a few of its rows. With mmap most of
it can stay on disk.
The complete per-tensor override map is in
Run Wladastic/Qwen3.8-Flash-Next-Ombro-v13b-GGUF with guIDE
Download guIDE — the AI-native code editor with local LLM inference and 69 built-in tools.
Source: Hugging Face · Compare models