QW

Qwen3.8 Flash Next MTP Strix Halo

QwenGGUF by EasiiXUpdated 2026-09-19
Download on Hugging Face
Parameters3.88B
Context256K
Layers49
Architectureqwen4exp
Smallest quant4.1 GB
Quick answer

Qwen3.8 Flash Next MTP Strix Halo needs about 5.5 GB of memory at Q8_0 (a 4.1 GB file) with an 8K context window, so it fits on a 6 GB GPU. This repository has 2 quants, from 4.1 GB (Q8_0) to 4.1 GB (Q8_0). Longer context needs more: at 32K tokens the KV cache alone adds about 3.3 GB.

Qwen3.8 Flash Next MTP Strix Halo – a GGUF sidecar for speculative decoding on AMD GPUs

This model is a GGUF sidecar of Qwen3.8‑Flash‑Next, tuned for AMD Strix Halo. It supports multi‑token prediction (MTP) and speculative decoding with a qwen4exp draft head. The model is 3.88 billion parameters and requires a llama.cpp build with the qwen4exp MTP draft head.

Good at

  • Fast speculative decoding on AMD Strix Halo
  • High token throughput in code generation
  • Efficient long‑context handling via QSA

Limits

  • Requires llama.cpp with qwen4exp MTP draft head
  • Speculative decoding needs --spec-draft-p-min 0.75
  • Approx. 5.5 GB VRAM at 8k context

Best for

  • AMD Strix Halo inference
  • High‑throughput code generation
  • Long‑context agent workloads

Who it’s for. Developers and researchers running llama.cpp on AMD GPUs who need efficient speculative decoding. It is ideal for those who want to leverage the MTP sidecar for high‑throughput, long‑context tasks.

Quants & memory

QuantFileEst. memoryvs VRAMResultDownload
Q8_04.1 GB 5.5 GB
Fits Get file →
Q8_04.1 GB 5.5 GB
Fits Get file →

The only available quant is Q8_0, file size 4.14 GB. It requires about 5.5 GB VRAM at 8k context. For lower VRAM budgets, no other quant is provided in this release.

What GPU do you need for Qwen3.8 Flash Next MTP Strix Halo?

These figures add the file size, an fp16 KV cache at an 8K context window (worked out from this model’s own header) and about 0.5 GB for the runtime. They are estimates, so leave a little headroom, and expect to use less context or a smaller quant if you are right at the limit.

  • 4 GB VRAM: no quant fits at an 8K context.
  • 6 GB VRAM: the best quant that fits is Q8_0 (4.1 GB file, about 5.5 GB in use).

Which Qwen3.8 Flash Next MTP Strix Halo quant should you choose?

Q8_0 (4.1 GB) is the middle option in this repository. Close to lossless. It costs roughly twice the memory of a 4-bit file, so it suits cases where you have plenty of VRAM.

New to quant names? Read GGUF quantization explained.

Specifications

Architecture from the GGUF header

Architectureqwen4exp
Parameters3.88B (approximate, mixture of experts)
Layers49
Trained context262,144 tokens
KV cache (fp16)98 KB per token

Source

Quantized byEasiiX
FamilyQwen
Licenseother
Downloads (HF)2,607
Last updated2026-09-19

Frequently asked questions about Qwen3.8 Flash Next MTP Strix Halo

How much VRAM does Qwen3.8 Flash Next MTP Strix Halo need?

About 5.5 GB at Q8_0 with an 8K context window. That is the 4.1 GB file, plus the KV cache and roughly 0.5 GB of runtime overhead. Smaller quants need less and longer context needs more.

Can I run Qwen3.8 Flash Next MTP Strix Halo on an 8 GB GPU?

Yes, with Q8_0: the 4.1 GB file needs about 5.5 GB at 8K context. Larger quants in this repository do not fit in 8 GB.

What is the file size of Qwen3.8 Flash Next MTP Strix Halo GGUF?

It depends on the quant: Q8_0 is 4.1 GB, Q8_0 is 4.1 GB.

What context length does Qwen3.8 Flash Next MTP Strix Halo support?

The GGUF header lists a trained context length of 262,144 tokens (256K). The KV cache grows with the context you use: about 0.8 GB at 8K and 3.3 GB at 32K.

Where can I download Qwen3.8 Flash Next MTP Strix Halo GGUF?

The files are on Hugging Face in the EasiiX/Qwen3.8-Flash-Next-MTP-Strix-Halo-GGUF repository. The download buttons on this page open the file on Hugging Face. ggufmodels.com does not host any model files.

What license does Qwen3.8 Flash Next MTP Strix Halo use?

Hugging Face lists the license as other. Licenses can differ between the quantized files and the original model, so check the repository and the base model before commercial use.

Can I run this model on 8 GB VRAM?

The Q8_0 quant requires about 5.5 GB VRAM at 8k context, so 8 GB is sufficient for inference but may limit batch size.

Does this model support vision tasks?

No vision support is provided; it is a text‑only model.

What build of llama.cpp is needed?

A llama.cpp build with the qwen4exp MTP draft head, such as the EngramHalo.cpp branch, is required.

Other GGUFs of Qwen3.8-Flash-Next

Similar models

Browse all →

More Qwen models

All Qwen →

Other small models

All small models →