Qwen3.8-Flash-Next-mixC-GGUF

Qwen3.8-Flash-Next, mixed-precision GGUF (map C)

A GGUF quantization of Qwen/Qwen3.8-Flash-Next, a 512-expert mixture-of-experts model with a 262 144-token context and image input. It averages 3.455 bits per weight, and on a 64 GB Apple silicon Mac all of its routed experts fit in memory at once.

• Routed experts: gate and up in IQ2_XXS, down in IQ4_NL, 40.4 GiB for all 24 576 experts.
• The per-layer n-gram table (26.8 GiB) is IQ4_NL and sits alone in part 2, which keeps it out of llama.cpp's GPU memory on a 64 GB Mac.
• The rest is Q5_K (token embedding, output), Q6_K and Q8_0.

Files

Put both parts in one folder and open part 1; the engine finds part 2 itself. The draft head is the model's own multi-token-prediction layer, with its experts quantized to IQ4_NL. It is only needed for speculative decoding.

Run it in llama.cpp

You need llama.cpp at commit 73a43d1f or later, the first with this architecture (qwen4exp).

llama-server -m Qwen3.8-Flash-Next-oQ2.0-mixC-00001-of-00002.gguf \
  --mmproj mmproj-Qwen3.8-Flash-Next-F16.gguf -c 131072 -ngl 99 -fa on --jinja

On a 64 GB Mac, a 131 072-token context with the projector fits with the default f16 cache. The full 262 144 does not: with f16 it fails at the first computation. It loads with -ctk q8_0 -ctv q8_0 (text only) or q4_0, but we have not measured what a quantized cache does to quality.

With the draft head (speculative decoding). This needs a llama.cpp build that contains pull request #28243. It was measured at 53b1389d. That build reads the draft layer from the model file itself, so first write a part 1 that contains it:

pip install gguf
python3 add-draft-head.py Qwen3.8-Flash-Next-oQ2.0-mixC-00001-of-00002.gguf mtp-head-IQ4_NL.gguf mixC-mtp
llama-server -m mixC-mtp-00001-of-00002.gguf -c 65536 -ngl 99 -fa on --jinja \
  --spec-type draft-mtp --spec-draft-n-max 2

The script copies part 1 with the head added (about 49 GB) and links part 2 under the new name. Keep the original parts: a llama.cpp without that pull request refuses the file with the head. Use a draft length of 1 or 2; longer was not faster. On a 64 GB Mac the head fits at a 65 536-token context with the projector, and not at 131 072.

Run it in Boxrunner

Boxrunner is a macOS app for local models (not yet publicly released). Choose Qwen3.8-Flash-Next 512x56B (ours, map C) in its model library. It downloads these files and sets context, cache and memory for your Mac. The draft head is off by default.

Measured speed

All measured on one Apple M5 Max with 64 GB (macOS 27), one request at a time, greedy, thinking off, f16 cache, no image loaded. Other machines will differ.

llama.cpp without the draft head, master 60b06ab9 (2026-09-19), 32 768-token context:

Both engines, with and without the draft head, 2026-09-26, 88 064-token context, generation in tokens/s. llama.cpp is PR #28243 at 53b1389d, Boxrunner is build adb8d70e3:

Boxrunner drafts one token per step. Its drafting numbers were taken after one long request. Right after start, drafting ran at about half that speed until the first long request.

No quality benchmark is reported here.

How it was made

Converted from Qwen/Qwen3.8-Flash-Next at revision de4b8e4d43b917e7706784d8bb445c9af86a3540 to Q8_0, then quantized with upstream llama.cpp 73a43d1f and a per-tensor type map. The importance matrix is AtomicChat's (its hash is in NOTICE). reproduce.sh records the commands.

Licence

The Qwen Community License 1.0, not Apache-2.0: see LICENSE and NOTICE. It allows copying, modifying and redistributing the weights, subject to two conditions. You must keep the notice, and very large commercial products must display the model name. Model-as-a-Service and AI Work Assistant businesses need a separate licence from Qwen for commercial use. Read LICENSE for the binding text.

Thanks to the Qwen team for the model and to the llama.cpp maintainers for the format and the tools. Neither endorses this repository.