Qwen3.8-Flash-Next · UD 3.1 IQ3_XXS GSQ-RCO Coder 320 PLE8
A 320-expert, code-focused derivative of ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-Coder-GGUF, now using Unsloth UD 3.1 UD-IQ3_XXS quantized weights while retaining multimodal capability.
This is a community derivative. It is neither an official Qwen release, nor an ISTA-DASLab release, nor an Unsloth release.
Overview
This model keeps the same 320-expert GSQ-RCO Coder selection used in the previous revision, but the retained quantized expert weights are now taken from Unsloth's UD 3.1 quants of Qwen3.8-Flash-Next, specifically the UD-IQ3_XXS variant. UD-IQ3_XXS is used here as the direct upgrade over the original ISTA IQ3_S quantization that was used in the previous version.
The model sits between ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-Coder-GGUF and the unpruned full 512-expert base quantization. Compared with the previous build, the expert set remains the same; only the source quantization for the retained tensors has changed from ISTA's IQ3_S to unsloth's UD-IQ3_XXS.
The 320-expert layout is designed to fit all retained experts and the 256k context within a roughly 45 GB memory budget under the Splash 1.5.0 runtime, with the n-gram/PLE table left on SSD.
Since the n-gram table does not need to reside in memory and its bandwidth requirements are low, this model ships with the Q8 PLE taken from Unsloth's UD-Q5_K_XL directory for Qwen3.8-Flash-Next. The shard-count metadata has been adjusted to fit the two-shard layout used in the ISTA-based releases. If desired, the PLE-containing second shard can be replaced with the corresponding smaller PLE table from ISTA's IQ3_S second shard or the identical pruned Coder release.
The quantized expert data is taken from the unpruned Unsloth UD-IQ3_XXS full release, then selected/pruned to the 320-expert Coder layout; no requantization was performed.
Although no extensive testing or evaluation has been performed, the model appears to work well for C++ and Python coding workflows in English.
Expert Selection
The baseline selection is the 256 experts from the v2 release of ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-Coder-GGUF. The remaining 64 experts were taken from ISTA's v1 initial release in the order determined in sh0wie/Qwen3.8-Flash-Next-REAP-288-MLX-4bit. If space remained for additional experts, they were added in saliency order from that list.
| blk | Coder v2 | Coder v1 | Remainder |
|---|---|---|---|
| 0 | 256 | 54 | 10 |
| 1 | 256 | 64 | -7 |
| 2 | 256 | 55 | 9 |
| 3 | 256 | 64 | -15 |
| 4 | 256 | 60 | 4 |
| 5 | 256 | 64 | -1 |
| 6 | 256 | 63 | 1 |
| 7 | 256 | 64 | -5 |
| 8 | 256 | 64 | -14 |
| 9 | 256 | 64 | -20 |
| 10 | 256 | 64 | -13 |
| 11 | 256 | 64 | -11 |
| 12 | 256 | 64 | -5 |
| 13 | 256 | 64 | -22 |
| 14 | 256 | 64 | -6 |
| 15 | 256 | 64 | -6 |
| 16 | 256 | 64 | -5 |
| 17 | 256 | 64 | -7 |
| 18 | 256 | 59 | 5 |
| 19 | 256 | 64 | -1 |
| 20 | 256 | 64 | -2 |
| 21 | 256 | 54 | 10 |
| 22 | 256 | 60 | 4 |
| 23 | 256 | 64 | -9 |
| 24 | 256 | 64 | -3 |
| 25 | 256 | 64 | -4 |
| 26 | 256 | 64 | -4 |
| 27 | 256 | 64 | 0 |
| 28 | 256 | 58 | 6 |
| 29 | 256 | 64 | -14 |
| 30 | 256 | 64 | -6 |
| 31 | 256 | 57 | 7 |
| 32 | 256 | 62 | 2 |
| 33 | 256 | 59 | 5 |
| 34 | 256 | 64 | -1 |
| 35 | 256 | 62 | 2 |
| 36 | 256 | 60 | 4 |
| 37 | 256 | 51 | 13 |
| 38 | 256 | 47 | 17 |
| 39 | 256 | 64 | 0 |
| 40 | 256 | 64 | -1 |
| 41 | 256 | 57 | 7 |
| 42 | 256 | 35 | 29 |
| 43 | 256 | 57 | 7 |
| 44 | 256 | 46 | 18 |
| 45 | 256 | 51 | 13 |
| 46 | 256 | 46 | 18 |
| 47 | 256 | 64 | 0 |
(A negative remainder means that not all distinct experts from the Coder v1 release could fit. Positive remainders were filled in saliency order.)
Pruning
Pruning was performed by applying prune_raw_gguf.py from webmp3/Sakura-Qwen3.8-Flash-Next-Overview to the UD-IQ3_XXS variant of Unsloth's full 512-expert Qwen3.8-Flash-Next GGUF release.
License
The weights inherit the license of the base model (Qwen3.8-Flash-Next).