kingjones777/Mistral-Small-4-119B-ROCmFP4-GGUF

⚠️ STOCK llama.cpp WILL NOT LOAD THIS MODEL

🚀 37.86 tok/s on a 119B MoE — 63.07 GiB, 7 GiB smaller than UD-Q4_K_XL.

Mistral-Small-4-119B-A6.5B — ROCmFP4 (tier 102 COHERENT) GGUF

A 4-bit ROCmFP4 quantization for AMD gfx1151 (Ryzen AI MAX+ 395 / Strix Halo), quantized from BF16 (222 GiB) — a lossless source, not a requantization of a lower-bit build.

File	Mistral-Small-4-119B-2603-Q4_0_ROCMFP4_COHERENT.gguf (sharded)
Size	63.07 GiB
BPW	4.55
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.

---

Files — where each build lives

build	ftype	size	where
4-bit COHERENT (2 shards)	102	63.07 GiB	in this repo
8-bit plain	111	114.38 GiB	in this repo · also at …-ROCmFPX-Q8_0-GGUF
8-bit AGENT	115	116.15 GiB	➡️ …-ROCmFPX-Q8_0-AGENT-GGUF

The AGENT build is hosted in its own repo rather than mirrored here — at ~116 GiB the duplication is not worth it, and (see the warning above) neither 8-bit build fits in GPU memory on a 128 GB Strix Halo. For a single Strix Halo, use the 4-bit build in this repo.

The separate per-quant repos exist so a user searching HF for a specific quant finds it directly.

All quant variants

variant	ftype	size	bpw	GPU on 128 GB Strix Halo	decode
4-bit COHERENT	102	63.07 GiB	4.55	✅ fits	37.86 tok/s
8-bit AGENT	115	116.15 GiB	8.39	⛔ does not fit	not measurable on this box
8-bit plain	111	114.38 GiB	8.26	⛔ does not fit	not measurable on this box

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

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	63.07 GiB	37.86	[37.83 – 38.65] (tight)
UD-Q4_K_XL	70 GiB	18.75	[15.99 – 34.30] (wide)

Ranges disjoint (34.30 < 37.83). The median ratio looks like +102%, but the baseline's own variance is large — we report the win without leaning on that headline number. Note the baseline is UD-Q4_K_XL, not Q4_K_M.

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 · shexp 108× Q8_0 (shared-expert protection) · router 36× F32 · norms 145× F32 · bulk TYPE_100 (288)

119B total / 6.5B active. Built with --tensor-type shexp=q8_0; shared experts are dense (they see every token) so their error is systematic, not averaged.

---

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.