Top-1 token agreement vs. unquantized Q8_0 reference, held-out corpus, ~2-bit class
unsloth UD-Q2_K_XL73.45 GiB
76.35 %
Agention AP-IQ2_S76.03 GiB
79.14 %

Our smallest build beats the leading community build in the same size class by 2.78 points of top-1 agreement, at roughly the same file size. Full methodology in the research writeup.

BuildFormatSizeTop-1 agreement*Target
AP-IQ2_S (smallest) GGUF, mainline-compatible 76.03 GiB 79.14% Any llama.cpp build, -ngl 99
AP-IQ3_XXS GGUF, mainline-compatible 80.72 GiB 81.85% Any llama.cpp build, -ngl 99
AP-IQ4_XS GGUF, mainline-compatible 84.24 GiB 82.89% Any llama.cpp build, -ngl 99
ROCmFP4-FAST-imatrix ROCmFPx, 4.23 bpw 87.06 GiB 84.56% Strix Halo (Radeon 8060S, 96 GB unified)
AP-Q4_K_XL GGUF, mainline-compatible 94.20 GiB 85.71% Any llama.cpp build, -ngl 99
AP-Q5_K_M GGUF, mainline-compatible 103.57 GiB 86.48% Any llama.cpp build, -ngl 99
AP-Q5_K_XL (recommended) GGUF, mainline-compatible 112.51 GiB 86.63% Any llama.cpp build, -ngl 99

*Top-1 agreement is how often a build picks the same token as an unquantized Q8_0 reference, measured on a held-out corpus the model hasn't seen. We moved off wikitext-2 perplexity because this model's per-layer n-gram table has memorized it; see the writeup for why, including for our own earlier numbers. "AP" is Agention Precision, a per-layer compression profile, distinct from the ROCmFPx format used in the Strix Halo build.

Size for size, we haven't found a mainline-compatible lineup for this model that beats these numbers on the same benchmark. If you have, we'd genuinely like to see it, and we'll measure it the same way we measured everything above.

Browse all builds on HuggingFace ↗

Size vs. quality across the lineup

The same relationship holds across every tier we publish: at each size where we have a matching unsloth tier, ours measures higher top-1 agreement. ROCmFP4-FAST, our fork-only build for Strix Halo, lands on the same curve as the mainline AP tiers rather than above it; its case for existing is speed, not extra quality-per-byte, see the research writeup for the full comparison.

Top-1 token agreement vs. unquantized Q8_0 reference, by file size, held-out corpus
Top-1 agreement vs. file size Nine quantized builds of Qwen3.8-Flash-Next, Agention AP, unsloth UD, and Agention's fork-only ROCmFP4-FAST, plotted by file size against top-1 token agreement with an unquantized Q8_0 reference on a held-out corpus. The AP curve sits above the UD curve at every comparable size; ROCmFP4-FAST lands on the AP curve. 75% 80% 85% 70 80 90 100 110 GiB Agention AP-IQ2_S · 76.03 GiB · 79.14% top-1 Agention AP-IQ2_S 76.03 GiB · 79.14% top-1 Agention AP-IQ3_XXS · 80.72 GiB · 81.85% top-1 Agention AP-IQ3_XXS 80.72 GiB · 81.85% top-1 Agention AP-IQ4_XS · 84.24 GiB · 82.89% top-1 Agention AP-IQ4_XS 84.24 GiB · 82.89% top-1 Agention AP-Q4_K_XL · 94.20 GiB · 85.71% top-1 Agention AP-Q4_K_XL 94.20 GiB · 85.71% top-1 Agention AP-Q5_K_M · 103.57 GiB · 86.48% top-1 Agention AP-Q5_K_M 103.57 GiB · 86.48% top-1 Agention AP-Q5_K_XL · 112.51 GiB · 86.63% top-1 Agention AP-Q5_K_XL 112.51 GiB · 86.63% top-1 unsloth UD-Q2_K_XL · 73.45 GiB · 76.35% top-1 unsloth UD-Q2_K_XL 73.45 GiB · 76.35% top-1 unsloth UD-IQ4_XS · 87.24 GiB · 83.40% top-1 unsloth UD-IQ4_XS 87.24 GiB · 83.40% top-1 Agention ROCmFP4-FAST (fork-only) · 87.06 GiB · 84.56% top-1 ROCmFP4-FAST (fork) 87.06 GiB · 84.56% top-1
Agention AP unsloth UD ROCmFP4-FAST (fork-only)

Every point measured on the same Q8_0 reference and the same held-out corpus. Hover or tap a point for exact figures; full numbers are in the table above.

Running these

The AP builds are stock GGUF: any current llama.cpp build loads the weights. Speculative decoding needs our fork, since MTP support for this model isn't upstream yet; the draft head adapts its length to measured acceptance instead of guessing a fixed depth. If you're VRAM-constrained, AP-IQ2_S runs in roughly two-thirds the footprint of AP-Q5_K_XL and still leads the field in its size class; swap the model path below and the same flags apply.

llama-server -m AP-Q5_K_XL/Qwen3.8-Flash-Next-AP-Q5_K_XL.gguf \
  -md Qwen3.8-Flash-Next-MTP-Q8_0.gguf \
  -ngl 99 --n-gpu-layers-draft 99 \
  --spec-type draft-mtp --spec-draft-adaptive \
  --spec-draft-n-min 2 --spec-draft-n-max 4

ROCmFP4-FAST needs our fork: it adds the ROCmFPx quant types and the per-head n-gram layout, plus the Vulkan fixes that keep throughput from collapsing at long context. On the same file, at 32k context, that's 2.75× the prefill and 1.9× the generation speed of stock llama.cpp; stock decays 55% in generation from 512 to 32k tokens, this fork only 10%. With the speculative MTP head on top, we've seen up to 40 tok/s generation on a Strix Halo.

git clone https://github.com/LaurentZuijdwijk/llama.cpp
cd llama.cpp && git checkout vulkan/qwen4exp-rocmfpx
cmake -B build -DGGML_VULKAN=ON -DCMAKE_BUILD_TYPE=Release
cmake --build build -j

./build/bin/llama-server \
  -m Qwen3.8-Flash-Next-ROCmFP4-FAST-v2-ple16.gguf \
  -ngl 99 -ctk q8_0 -ctv q8_0 -fa on

Full setup, vision, and speculative-decoding instructions are in each repo's README on HuggingFace.

View the fork on GitHub ↗