kingjones777/Mellum2-12B-A2.5B-Thinking-ROCmFP4-GGUF

🔧 Runtime: build the ROCmFPX fork below

Stock llama.cpp will not load this file. You need both the mellum architecture and the ROCmFP4 tensor types in one tree. Upstream charlie12345/ROCmFPX has the ROCmFP4 types but not mellum. Our fork has both:

kingjones30/ROCmFPX — a fork of charlie12345/ROCmFPX, branch main.

git clone https://github.com/kingjones30/ROCmFPX.git
cd ROCmFPX
cmake -B build -DGGML_HIP=ON -DGPU_TARGETS=gfx1151 -DGGML_NATIVE=ON -DCMAKE_BUILD_TYPE=Release
cmake --build build --target llama-server llama-quantize -j$(nproc)

Verified 2026-08-27 on gfx1151: clean clone → 0 build errors → llama-server loads a mellum ROCmFP4 GGUF from this family and generates coherent text.

⚠️ STOCK llama.cpp WILL NOT LOAD THIS MODEL

The Mellum architecture is not merged upstream (PR #23966) — the patch is in patches/ in this repo.

🚀 104.99 tok/s — +6.9% faster than Q4_K_M, ranges disjoint.

Mellum2-12B-A2.5B-Thinking — ROCmFP4 (tier 102 COHERENT) GGUF

A 4-bit ROCmFP4 quantization for AMD gfx1151 (Ryzen AI MAX+ 395 / Strix Halo), quantized from BF16 GGUF (22.6 GiB) — a lossless source, not a requantization of a lower-bit build.

File	Mellum2-12B-A2.5B-Thinking-Q4_0_ROCMFP4_COHERENT.gguf
Size	6.4907 GiB
BPW	4.59
ftype	Q4_0_ROCMFP4_COHERENT (102)

⛔ Requires a llama.cpp with the ROCmFP4 quant types

Q4_0_ROCMFP4_COHERENT (ftype 102) exists only in charlie12345/ROCmFPX, not upstream llama.cpp. Ignore the auto-generated "Use this model" commands above.

---

All quant variants

All measured on one box, one binary (Ryzen AI MAX+ 395, gfx1151, ROCm 7.2.4), median of 3, warm-up discarded — so these rows are directly comparable.

variant	ftype	size	bpw	decode (median)	range
4-bit COHERENT	102	6.49 GiB	4.59	104.99	104.96 – 105.73
8-bit AGENT	115	11.88 GiB	8.39	74.41	74.36 – 74.42
8-bit plain	111	11.70 GiB	8.27	75.72	75.72 – 75.73

Repos: 4-bit · 8-bit AGENT · 8-bit plain

⚠️ AGENT is SLIGHTLY SLOWER here — 74.41 vs 75.72. The opposite of the Instruct variant, which tells you the difference is model-specific and small. Neither 8-bit build approaches the 4-bit build's 104.99 tok/s.

On AGENT generally: it keeps more tensors at true Q8_0 instead of the packed 8-bit type. That raises MTP draft acceptance on models which have an MTP head (measured +6.2% on Qwen3.8-27B). Mellum2 has no MTP head, so there is nothing for the extra precision to feed and the two 8-bit builds differ only marginally — in either direction.

Measured

Ryzen AI MAX+ 395 (gfx1151, 128 GB unified, ROCm 7.2.4). Median of 3+, warm-up discarded, otherwise-idle box. Correctness at the model's official sampling.

build	size	decode (median)	range
this build	6.4907 GiB	104.99	[104.96 – 105.73]
Q4_K_M	—	98.17	[98.16 – 98.87]

+6.9%, ranges disjoint — a real difference, not noise.

Correctness: 17×23 ⇒ ✅ 391 · capital of Japan ⇒ ✅ Tokyo · days in 2024 ⇒ ✅ 366

Per-tensor types (audited in the finished file)

output.weight Q6_K · token_embd Q6_K · router ffn_gate_inp F32 · norms F32 · 339 tensors

tie_word_embeddings is false, so both --output-tensor-type and --token-embedding-type apply.

---

What was NOT measured

• No perplexity run, and no quality A/B against the baseline or the source. The checks above are memorized-fact prompts — necessary but not sufficient; a damaged model can pass them.
• No long-context testing.
• No tool-calling evaluation.

Base model licence inherited; credit for the model goes to its authors.