gilde

← All models

Qwen 3.8 Flash Next (180B-A6B, IQ3_XXS)

generalQwen Community 1.0tools

Qwen's experimental preview of the architecture behind Qwen4: a 125B-parameter MoE with 6B active per token, plus a 51B n-gram embedding table designed for cheap offload. This Unsloth UD-IQ3_XXS build is an 82 GB split GGUF whose 26.8 GiB n-gram table can be demand-paged; it is the practical variant for 64 GB system RAM plus a 32 GB-class discrete GPU. Gezel starts it conservatively at 64K context and one slot, then chooses the expert GPU/CPU split from the machine's measured VRAM.

At a glance

Parameters
180B
Quantization
UD-IQ3_XXS
Context window
262K tokens
Approx. size
82.0 GB
Engines
llama.cpp
License
qwen-community-1.0 (custom-restricted)
Version
v1.0.0 , released 2026-09-06
Family
qwen

Tuning

Defaults a gezel client applies out of the box, from real evaluation runs against this exact quantization.

sampling.temperature
1
sampling.topP
0.95
sampling.topK
20
sampling.minP
0
sampling.maxTokens
12288
sampling.repetitionPenalty
1
sampling.presencePenalty
0
reasoning.thinkingBudget
4096
reasoning.enableThinking
true
reasoning.templateKwargs
{"reasoning_effort":"xhigh"}
profiles
thinking-general, thinking-deep, thinking-coding, thinking-precise, instruct, creative

Model behaviors

Client-side behavior modules the catalog enables for this model (reasoning-tag handling, fabrication detection, prompt shaping).

Sources

llama.cpp (GGUF)
unsloth/Qwen3.8-Flash-Next-GGUF · UD-IQ3_XXS
Upstream
https://huggingface.co/Qwen/Qwen3.8-Flash-Next
Manifest
View on GitHub

Tags