Qwen3.6-14B-A3B-vibetuned ROCmFPX STRIX_LEAN — GGUF

ROCmFPX Q4_0_ROCMFP4_STRIX_LEAN quant of tvall43/Qwen3.6-35B-A3B-Heretic, down-pruned to a 14B-A3B variant by the vibetuned author.

Built with charlie12345/ROCmFPX on a Radeon RX 9060 XT 16 GB (gfx1200), ROCm 7.2.3, NixOS 25.11. Quantized 2026-06-27 with build commit 11d76c2.

This is not a stock llama.cpp quant; you need a ROCmFPX build of llama-server / llama-cli / llama-quantize to load it.

Scope of these benchmarks — read this first

These numbers are a light baseline, not a thorough ROCmFPX evaluation. The mesh's bench framework is built for production agent workload regression-detection on the local stack, not for the kind of multi-axis sweep that upstream quant maintainers typically publish. Specifically:

• Harness scope is bounded. The numbers below come from the mesh's mesh_eval (6 tests, 4 deterministic + throughput) + hermes_loop_eval (5 agent scenarios) + a ctx_scaling test at 4 K → 32 K (the 64 K and 128 K ctx requests returned HTTP 400 from this server config — see "What's NOT in this repo"). That's a regression suite, not a quality benchmark.
• Sample sizes are small. Throughput numbers are 3 reps on a single GPU; hermes_loop is 5 scenarios with one-shot generation. None are powered for statistical significance on a per-token level.
• No perplexity / wikitext / MMLU / GSM8K. The mesh's stack isn't a quality benchmark — those are upstream ROCmFPX's territory. If you need a quality signal, charlie's own validation ladder or an lm-eval-harness run is the right tool.
• Single GPU class. All measurements are on a 16 GB RDNA4 (RX 9060 XT, gfx1200). No Strix unified-memory, no CDNA, no multi-GPU, no Vulkan, no CUDA. Cross-hardware generalization is not implied.
• No human eval. "Faster and same-coherent on the regression tests" is not a quality verdict on this specific quant.

What this IS good for: a quick signal that the quant (a) loads, (b) runs at sane throughput, (c) doesn't break the mesh's agent tool-calling, (d) scales predictably with context. What this is NOT good for: claiming "this is the best quant of this model," reproducing academic benchmark results, or substituting for upstream's validation work.

For a rigorous view, the parent repo tvall43/Qwen3.6-35B-A3B-Heretic and the upstream Qwen3.6 GGUF variants (e.g. on bartowski/) are the place to look.

What we measured

Hardware: Node B, AMD Ryzen 9 5900XT 16-core, Radeon RX 9060 XT 16 GB (gfx1200), ROCm 7.2.3, NixOS 25.11 Software: charlie12345/ROCmFPX main @ 11d76c2 Source GGUF: Qwen3.6-14B-A3B-vibetuned-F16.gguf (F16, 26 GB) from the vibetuned author's published artifact Same-stack comparison: Q6_0_ROCMFPX (6-bit ROCmFPX, 11 GB file) on the same source

Agent-loop throughput — STRIX_LEAN vs Q6_0_ROCMFPX (hermes_loop, same harness, same source)

Both quants pass all 5 scenarios. The 4-bit STRIX_LEAN is 2-2.5× faster than the 6-bit Q6_0 on this MoE arch, and 36 % smaller on disk (7.0 GB vs 11 GB). This is the headline finding for this model.

mesh_eval (raw JSON: raw-mesh-eval-vibetuned-14b-strix-lean.json)

hermes_loop (raw JSON: raw-hermes-loop-vibetuned-14b-strix-lean.json)

Context scaling (raw JSONs: ctx-scaling-vibetuned-strix-lean-64k-*.json, ctx-scaling-vibetuned-strix-lean-128k-*.json)

Findings:

• Decode throughput holds at 50 t/s across 4 K → 64 K ctx (single-batch). The 4-tick spread vs Ornith's 48 t/s is within harness noise.
• Prompt-processing decline from 4 K → 32 K is ~38 %, consistent with KV-cache pressure scaling.
• This server's 128 K cap is a config limit, not a model limit — the parent Qwen3.6 35B-A3B has 256 K native ctx, and the smaller 14B-A3B variant should fit 128 K on a 24+ GB card.

KV cache type sweep (extrapolated from Ornith, head_dim=128)

The mesh's KV-type sweep was run on Ornith 9B (also head_dim=128). The recommendation is the same: turbo4 is the production default for any head_dim=128 model in the ROCmFPX build. See the Ornith 9B ROCmFPX STRIX_LEAN repo for the full sweep data.

Quick start

# Build llama.cpp with ROCmFPX
git clone https://github.com/charlie12345/ROCmFPX
cd ROCmFPX
cmake -S . -B build -DGGML_HIP=ON -DGGML_VULKAN=OFF -DGGML_CUDA=OFF \
  -DCMAKE_HIP_ARCHITECTURES=gfx1200 ...
cmake --build build --target llama-server llama-cli llama-quantize

# Serve (131 072 ctx, turbo4 KV for head_dim=128, fa=on)
./build/bin/llama-server \
  -m Qwen3.6-14B-A3B-vibetuned-ROCmFPX-STRIX_LEAN.gguf \
  -np 1 -c 131072 \
  -ctk turbo4 -ctv turbo4 \
  -kvo -cram 32768 -fa on

Reproduce the quant

SRC=/path/to/Qwen3.6-14B-A3B-vibetuned-F16.gguf

~/ROCmFPX/build-rdna4/bin/llama-quantize \
  "$SRC" \
  Qwen3.6-14B-A3B-vibetuned-ROCmFPX-STRIX_LEAN.gguf \
  Q4_0_ROCMFP4_STRIX_LEAN

Quantize time: ~2-4 min warm-cache, CPU-only.

Files in this repo

What's NOT in this repo (caveats)

• Stock llama.cpp will not load this file. The ROCmFP4 weight format is unique to charlie12345/ROCmFPX.
• No CUDA / non-AMD GPU bench. All measurements are RDNA4 (gfx1200). Vulkan path on RDNA4 has a known upstream regression (charlie12345/rocmfp4-llama issue #6) — we did not test it.
• 128 K ctx is HTTP 400 on this server. The parent Qwen3.6 35B-A3B has 256 K native ctx; the 14B-A3B vibetuned variant should fit 128 K on a 24+ GB card. We tested up to 64 K successfully; the 128 K failure is the server's --ctx-size cap, not a model limit.
• The vibetuned author is tvall43 (per the GGUF metadata general.base_model.0.organization); the Qwen3.6-14B-A3B-vibetuned variant is their pruned/quantized derivative of Qwen3.6-35B-A3B-Heretic. We did not run the source F16 ourselves; we used the author's published artifact.
• No MTP / speculative-decode bench on this file. The mesh's MTP head work was done on a different Qwen3.6 family model. MTP draft heads for this variant are not packaged.
• No vision/multimodal test. This variant is text-only.
• KV cache sweep was run on Ornith 9B (same head_dim=128, same arch family), not on this file directly. The recommendation transfers but the specific VRAM/t-s numbers are from Ornith.

Provenance

• Source model: tvall43/Qwen3.6-35B-A3B-Heretic — the 14B-A3B vibetuned variant is derived from this 35B MoE
• Source model license: apache-2.0
• Quantizer: charlie12345/ROCmFPX main @ 11d76c2 (2026-06-27)
• Quantizer license: MIT
• Build hardware: Node B, AMD Ryzen 9 5900XT 16-core, Radeon RX 9060 XT 16 GB (gfx1200), ROCm 7.2.3, NixOS 25.11
• Build tooling: NixOS 25.11, ROCm store paths dynamic-discovered. See the meshina repo's references/nixos-rocm-external-build-recipe.md for the build env setup.
• Bench harnesses: scripts/mesh-bench/mesh_eval.py + scripts/mesh-bench/hermes_loop_eval.py + scripts/mesh-bench/ctx_scaling_bench.py from the meshina repo (private)
• Original bench report: raw/benchmarks/2026-06-27-rocmfpx-validation/briefs/2026-06-27-rocmfpx-rdna4-16gb.md in the meshina repo

License

• The Qwen3.6 vibetuned parent is apache-2.0 (per its HF model card).
• The charlie12345/ROCmFPX quantizer is MIT.
• The GGUF in this repo is a derivative of the apache-2.0 parent, produced with the MIT-licensed quantizer. Both upstream licenses are preserved.