Qwen3.8 Flash Next
Qwen's mixture-of-experts model with 125B parameters and only 6B active per token. Fast decoding for its size, if you have the memory to hold it.
Overview
Qwen3.8 Flash Next is a sparse mixture-of-experts model: 512 experts with 10 routed plus 1 shared active per token, about 6B activated parameters out of 125B (plus n-gram embeddings and an MTP head). Because so few parameters are active, it decodes much faster than a dense model of the same total size.
Hardware fit
Even at UD-Q4_K_XL the weights are 111 GB, so plan for a single 141 GB H200 or two 80 GB GPUs. UD-Q2_K_XL at 80 GB fits one H100 with a small context.
| Quantization | Weights |
|---|---|
| UD-Q2_K_XL | 79.6 GB |
| UD-Q4_K_XL | 111.3 GB |
| Q8_0 | 199.3 GB |
Sizes are the GGUF file sizes from unsloth/Qwen3.8-Flash-Next-GGUF. Budget extra VRAM for the KV cache, which grows with context size and parallel requests.
Deploying on Loadcore
- In the console, open Models and search for
unsloth/Qwen3.8-Flash-Next-GGUF. - Pick the quantization that fits the target server and start the download.
- Start an instance with the context size you need and GPU layers set to offload the full model.
- Point your tools at
http://your-server:9090/v1with your personal key.
Recommended settings
| Parameter | Value |
|---|---|
| temperature | 1.0 |
| top_p | 0.95 |
| top_k | 20 |
| repeat_penalty | 1.0 |
| min_p | 0.0 |
Same sampling guidance as the rest of the Qwen3.8 family. Instruct mode uses temperature 0.7, top_p 0.8 and presence_penalty 1.5. Check the license before commercial use, since it is not Apache 2.0.
Source: Qwen/Qwen3.8-Flash-Next.