Qwythos-9B-Claude-Mythos-5-1M-MTP ROCmFP4 COHERENT — GGUF

ROCmFP4 COHERENT quant of Qwythos-9B-Claude-Mythos-5-1M-MTP (Qwen3.5-9B dense, 1M YaRN, vision). Q6_K embeddings preserve the shared-embedding MTP head quality — 4/4 mesh_eval, 5/5 hermes_loop, +37.5% MTP throughput on a single 16 GB RDNA4 card. Built with charlie12345/ROCmFPX 11d76c2 for AMD ROCm (gfx1200).

⚠ This is NOT a stock llama.cpp quant. ROCmFP4 weight formats (Q4_0_ROCMFP4_COHERENT, Q4_0_ROCMFP4_STRIX_LEAN, Q8_0_ROCMFPX_AGENT) are unique to the charlie12345/ROCmFPX fork. Stock llama.cpp will exit with unknown quantization at load time. Use the ROCmFPX fork's llama-server / llama-cli.

Scope of these benchmarks — read this first

These numbers are a light baseline, not a thorough ROCmFP4 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 (4 deterministic tests + throughput) + hermes_loop_eval (5 agent scenarios). That's a regression suite, not a quality benchmark — it answers "does this quant still serve the mesh's agent stack correctly," not "is this the best possible ROCmFP4 quant of this model."
• 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, charlie12345'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 empero-ai/Qwythos-9B-Claude-Mythos-5-1M and the upstream Qwen3.5-9B are the place to look.

What we measured

ROCmFP4 COHERENT vs STRIX_LEAN vs AGENT (Node B, 16 GB RDNA4)

COHERENT is the winner. STRIX_LEAN's Q5_K embeddings cause thinking-leak on this shared-embedding MTP model. AGENT (Q8_0_ROCMFPX) is 9.1 GiB — too large for MTP compute buffers on 16 GB at any context size.

Cross-GPU: ROCm vs CUDA

CUDA Q6_K is faster per-token, but ROCm COHERENT delivers the same quality scores at 33% smaller footprint — critical for multi-model GPU slots.

mesh_eval — COHERENT MTP-OFF

Overall: 4/4 (vision excluded from core count)

hermes_loop — COHERENT MTP-ON (all 5/5)

MTP speedup context — first real MTP win on the mesh

This is the first model on the Sovereign Machina mesh where MTP speculative decoding delivers a large, reliable speedup (+37.5% on ROCm, +81.2% on CUDA Q6_K per the companion bench). Prior models with native single-file MTP (Ornstein-9B-v2) showed only +0-2% within noise. The difference: Qwythos is a dense Qwen3.5-9B with a more capable MTP head, where draft acceptance rates are high enough to recover the overhead.

Quick start

# ROCmFPX fork (required — stock llama.cpp won't load this quant)
git clone https://github.com/charlie12345/ROCmFPX
cd ROCmFPX
mkdir build && cd build
cmake -DLLAMA_HIPBLAS=ON -DAMDGPU_TARGETS=gfx1200 ..
make -j$(nproc)

# Serve with all production flags
HSA_OVERRIDE_GFX_VERSION=12.0.0 \
LD_LIBRARY_PATH=$(pwd)/bin \
./bin/llama-server \
  -m Qwythos-9B-Claude-Mythos-5-1M-MTP-ROCmFP4-COHERENT.gguf \
  --mmproj mmproj-Qwythos-9B-Claude-Mythos-5-1M-f16.gguf \
  --host 0.0.0.0 --port 8081 \
  -ngl 99 -c 131072 -np 1 -ub 2048 \
  --cache-type-k turbo4 --cache-type-v turbo4 --cache-ram 32768 --kv-unified \
  -fa on --metrics --fit off \
  --spec-type draft-mtp

Flags explained: -ub 2048 is required for Qwen3.5 hybrid SSM architecture. --fit off is required at 64K+ context on RDNA4 (Pattern 16 per ROCmFPX docs). turbo4 KV cache is free performance for head_dim=128 models (Qwen family). --spec-type draft-mtp enables native single-file MTP speculative decoding.

Reproduce the quant

~/ROCmFPX/build-rdna4/bin/llama-quantize \
  --allow-requantize \
  /path/to/Qwythos-9B-Claude-Mythos-5-1M-MTP-BF16.gguf \
  Q4_0_ROCMFP4_COHERENT \
  Qwythos-9B-Claude-Mythos-5-1M-MTP-ROCmFP4-COHERENT.gguf

Source quant: empero-ai/Qwythos-9B-Claude-Mythos-5-1M-GGUF → BF16 source → COHERENT re-quant.

Files in this repo

What's NOT in this repo (caveats)

• Stock llama.cpp will not load this file. The Q4_0_ROCMFP4_COHERENT weight format is unique to charlie12345/ROCmFPX. Use that fork's llama-server/llama-cli/llama-quantize.
• No CUDA / non-AMD GPU bench. All measurements are RDNA4 (gfx1200). Vulkan path on RDNA4 has a known upstream regression — we did not test it.
• 131K ctx is HTTP 400 if --fit off is omitted. Pattern 16 of the ROCmFPX docs: RDNA4 with HSA requires --fit off at 64K+ context to avoid HSA SEGV in hsaKmtWaitOnMultipleEvents_ExtCtx.
• Vision test failed on the ROCmFPX build. The ROCmFPX fork's mmproj path may not fully support vision encoding for this model on RDNA4. The parent model supports vision on CUDA/stock llama.cpp.
• No MTP bench on the AGENT quant. AGENT (9.1 GiB) cannot run MTP on 16 GB — compute buffers OOM regardless of context size.
• No quality benchmark (perplexity, MMLU, GSM8K). See "Scope of these benchmarks" above — this quant is validated for agent stack regression, not academic reproduction.
• Uncensored test result is equivocal. The eval returned empty output on a sensitive prompt — not a refusal, but not a clear pass either. No impact on agent tool-calling or coding benchmarks.
• The source is from empero-ai/Qwythos-9B-Claude-Mythos-5-1M safetensors, quantized via empero-ai/Qwythos-9B-Claude-Mythos-5-1M-GGUF BF16 intermediate. The chain is: Qwen/Qwen3.5-9B (Apache 2.0) → empero-ai (fine-tune) → BF16 GGUF → our COHERENT quant.
• 4.5+ GiB minimum VRAM. Doesn't fit on smaller cards. The mesh's 16 GB card runs it with ~11 GB headroom at 4K context, ~1 GB headroom at 131K.

Provenance

• Date: 2026-06-29/30
• Source model: empero-ai/Qwythos-9B-Claude-Mythos-5-1M (Apache 2.0, Qwen3.5-9B fine-tune with 1M YaRN RoPE + MTP)
• Intermediate GGUF: empero-ai/Qwythos-9B-Claude-Mythos-5-1M-GGUF (BF16 source)
• Quantizer: charlie12345/ROCmFPX 11d76c2 via llama-quantize --allow-requantize
• Build hardware: Node B — AMD Ryzen 9 5900XT 16C/24T, 64GB DDR4, AMD RX 9060 XT 16GB (gfx1200, RDNA4, ROCm 7.8)
• Build tooling: ROCmFPX fork with -DLLAMA_HIPBLAS=ON -DAMDGPU_TARGETS=gfx1200
• Bench harnesses: mesh_eval.py (6 tests), hermes_loop_eval.py (5 agent scenarios)
• Original bench report: raw/benchmarks/2026-06-30-qwythos-9b-rocm-bench/

License

• Parent model (empero-ai/Qwythos-9B-Claude-Mythos-5-1M): Apache 2.0
• Upstream base (Qwen/Qwen3.5-9B): Apache 2.0
• Quantizer (charlie12345/ROCmFPX): MIT
• This quant: Apache 2.0 (derived from Apache 2.0 parent)

This is a derivative quantized file. The license terms of the parent model apply to the use of this file. Verify commercial use terms with the parent model card before deployment.