Home›Models›qwen3.8-flash-next
qwen3.8-flash-next
Architecture mixture of experts48 LayersLicense qwen-community-1.0Maximum context 256kVocabulary 248,320
At IQ4_XS it is an 87.25 GiB file, and it fits on 4 of the 18 machines we measure (the one with the least memory: PC with Ryzen AI Max+ 395 and 128 GB unified). Here the file is what fills the memory: even at its maximum context of 256k tokens, the context adds 6.00 GiB — and the same content in German takes ×1.31 the tokens of English.
Model size at IQ4_XS
Measured by Local AI Scope
Context memory at 8k tokens
Measured by Local AI Scope
What German costs versus English
Measured by Local AI Scope
Sizes of the real published files
These are the sizes of the files actually published by the author, added up when the model ships split into parts. Not an estimate from the parameter count.
| Compression | Model size |
|---|---|
Q8_0 |
175.30 GiB |
IQ4_XS |
87.25 GiB |
This model also reads images. For that it needs a second file, the vision projector (mmproj-F16.gguf, 0.84 GiB), published next to the model and not included in the figures on this page: they count text only. If you are going to send it images, add it to the model size.
What changes between Q8_0, Q4_K_M and Q3_K_M (size, speed and quality) is something we measured ourselves in our GGUF quantization guide.
Context memory
Computed from the attention pattern this model declares layer by layer, not from the generic formula. That is why the figure is often far smaller than other calculators tell you. Besides that memory, this model keeps a second, smaller cache for its sparse-attention indexer: about 3 KiB per token, 0.02 GiB at 8k and 0.75 GiB at its maximum context. It is derived from the code and the file header, not declared by the author, and it is not added to the figures on this page. Added in, it changes none of the verdicts in the machine table.
- 8k tokens → 0.19 GiB
- 32k tokens → 0.75 GiB
- 128k tokens → 3.00 GiB
What it costs in each language
Measured by running this model’s own tokenizer over the same content in four languages. Tokens are what you pay for in context, in memory and in your cloud bill.
| Kind of text | EN | ES | FR | DE |
|---|---|---|---|---|
| Code | — | ×1.20 | ×1.30 | ×1.46 |
| Contracts | — | ×1.16 | ×1.23 | ×1.31 |
| Business email | — | ×1.12 | ×1.22 | ×1.17 |
| Support tickets | — | ×1.24 | ×1.40 | ×1.25 |
English is the baseline: every figure is how many times more tokens the same content costs in that language. Average across the whole corpus: DE ×1.31.
Where it fits
Fits means the model, its context memory and the working headroom all fit at the stated context, leaving the system its share. 4 of the 18 machine profiles tested (IQ4_XS, 8k tokens).
One piece of this model behaves differently depending on the program that runs it. 26.82 GiB of the IQ4_XS file are an n-gram table (per_layer_token_embd) that is only ever read a few rows at a time. llama.cpp v0.5.0 and later leave it on disk by default and read those rows on demand, so what has to fit in memory drops to 60.43 GiB. On that basis it would also fit on Xiaomi AI Cube, 80 GB — engineering prototype, up to 256k tokens of context. Closest miss: Mac with M4 Max and 64 GB · Mac mini with M5 Pro and 64 GB, short by 0.20 GiB. The table above counts the whole file, because that holds for every program; the saving is checked for llama.cpp v0.5.0, in its code and in the file header, and for Ollama’s MLX engine on Mac from v0.40.16 on (read in its code on 11 November 2026), but only with its NVFP4 download (qwen3.8-flash-next:125b-a6b-nvfp4, 105 GB in Ollama’s library): with any other quantization that engine loads the table like the rest. The code of llama.cpp v0.5.0 marks its support for this architecture as provisional, pending a full reimplementation (read on 29 September 2026).
