K2-Horizon-32B-ROCmFP4-GGUF

K2-Horizon-32B — ROCmFP4 for AMD Strix Halo (gfx1151)

Four ROCmFP4/ROCmFPX tiers of IFM/K2-Horizon-32B, quantised from the official BF16 GGUF and measured on a Ryzen AI Max+ 395 (Radeon 8060S, gfx1151, 128 GB unified, ROCm 6.4.1). K2-Horizon-32B is a dense k2_horizon decoder — 64 layers, hidden 5120, 64 attention heads, rope_head_dim 128 — and a reasoning model.

Every number on this card was measured on these exact files. Nothing is estimated.

Which file should I use?

Start with FAST unless you specifically want 8-bit weights.

Quick start — the backend is one flag

# HIP / ROCm — best prefill (long prompts, RAG, agentic)
llama-server -m K2-Horizon-32B-Q4_0_ROCMFP4_FAST.gguf \
  -dev ROCm0 -fa on -ngl 999 --no-mmap -fit off -np 1 \
  -b 2048 -ub 1024 -t 16 --poll 100 -c 8192 --jinja

# Vulkan — SAME FILE. Only -dev changes.
llama-server -m K2-Horizon-32B-Q4_0_ROCMFP4_FAST.gguf -dev Vulkan0 ...

⛔ You need a build with BOTH k2_horizon and the ROCmFP4 types

These files declare general.architecture = k2-horizon and use the ROCmFP4/ROCmFPX tensor types (ggml types 100-119). Two independent pieces of support are required, and no single public build has both yet:

• k2_horizon architecture — lives in the MBZUAI-IFM llama.cpp fork; an upstream PR is in progress.
• ROCmFP4/ROCmFPX tensor types — from ROCmFPX.

This repository ships k2-horizon-on-rocmfpx.patch — the k2_horizon arch applied on top of a ROCmFPX base, exactly the tree these files were built and verified against. Build:

git clone https://github.com/ROCmFPX/ROCmFPX.git && cd ROCmFPX
curl -fLO https://huggingface.co/kingjones777/K2-Horizon-32B-ROCmFP4-GGUF/resolve/main/k2-horizon-on-rocmfpx.patch
git apply k2-horizon-on-rocmfpx.patch
cmake -B build -DGGML_HIP=ON -DGGML_VULKAN=ON -DAMDGPU_TARGETS=gfx1151 -DGGML_NATIVE=ON -DCMAKE_BUILD_TYPE=Release
cmake --build build -j$(nproc)

⛔ Serving note — this is a reasoner

K2-Horizon-32B emits chain-of-thought into reasoning_content before it answers. A short token budget returns HTTP 200 with an empty content and all the tokens in reasoning — that is the model thinking, not a broken file. Give it room: tool calls and short answers are fine at ~500 tokens; long-form may need more. Verified: at 40 tokens the content was empty; at 500 it answered correctly.

Head handling

Quantised at each tier's native routing with no forced --output-tensor-type override. On this family the unprotected head performs well; the tiers are posted as-built.

Reproduction block

model     : IFM/K2-Horizon-32B  (official BF16 GGUF, K2-Horizon-32B-BF16.gguf, 64.8 GB)
            arch k2-horizon (dense): 64 layers, hidden 5120, 64 heads, rope_head_dim 128
port      : MBZUAI-IFM/llama.cpp @ model/K2Horizon (35999d1) k2_horizon arch
            applied onto ROCmFPX fork @ 85d8f7e  (patch shipped in this repo)
quantize  : llama-quantize  <in> <out> <ftype>   (no head-type override)
            Q4_0_ROCMFP4_COHERENT (102) · Q4_0_ROCMFP4_FAST (103)
            Q4_0_ROCMFP4_STRIX_LEAN (106) · Q8_0_ROCMFPX_AGENT (115)
serve     : llama-server -dev {ROCm0|Vulkan0} -fa on -ngl 999 --no-mmap -fit off
            -np 1 -b 2048 -ub 1024 -t 16 --poll 100 -c 8192 --jinja
box       : Ryzen AI Max+ 395 / Radeon 8060S (gfx1151) / 128 GB unified, ROCm 6.4.1
verified  : loads on GPU, generates correct output ("capital of France" -> "Paris")
measured  : 2026-09-04

SHA256

98f213c632cbbbffecc87a136603cc5088c3586ff6fdf15a72427b622a6480c6  K2-Horizon-32B-Q4_0_ROCMFP4_COHERENT.gguf
35baebe33a61147ba8f074a9b1e24f25c5936add7675f755b6f1866f1af4e9a8  K2-Horizon-32B-Q4_0_ROCMFP4_FAST.gguf
b737899d4f7f013549bad51002b4b6feee4b7c1a48c6eefe6b8d54ca3019dd53  K2-Horizon-32B-Q4_0_ROCMFP4_STRIX_LEAN.gguf
64c7c420a197bb7f96de1afafeb0821829e64e8a9b361520875c8ec56734ec6d  K2-Horizon-32B-Q8_0_ROCMFPX_AGENT.gguf

Credits

IFM / MBZUAI — the K2-Horizon-32B base model (Apache-2.0) and the k2_horizon llama.cpp architecture. ROCmFP4/ROCmFPX quantisation types are the work of ROCmFPX. This repository adds only the ROCmFP4 quant ladder, the arch-on-ROCmFPX patch, and the measurements.