jcbtc/chadrock-35b-ace-saber-rocmfp4-mtp

Chadrock-35B Ace Saber ROCmFP4 MTP

Chadrock-35B Ace Saber is a ROCmFP4/MTP GGUF for AMD Ryzen AI Max+ 395 / Strix Halo systems.

The model behavior comes from the Ace Saber build by @DJLougen. The runtime speed comes from charlie12345, also known as @Italianclownz on X/Twitter, and his custom ROCmFP4 llama.cpp fork.

This GGUF will not run correctly with stock llama.cpp. You need the custom charlie12345/rocmfp4-llama build because this file uses ROCmFP4 tensor types that upstream llama.cpp does not currently understand.

The model file is already provided here. You do not need to rebuild or quantize the model.

Why This Mix

Ace Saber gives the model its coding, agentic, and tool-use behavior. Chadrock/ROCmFP4 gives it the speed profile needed to feel good locally on AMD unified-memory hardware.

The goal is not just another Qwen3.6 quant. The goal is:

• Ace Saber behavior from @DJLougen
• Qwen3.6 35B-A3B MoE efficiency
• MTP speculative decoding
• ROCmFP4 tensor-aware quantization
• high-throughput local serving on Ryzen AI Max+ 395 / Strix Halo

Technical Metadata

Hugging Face may round the parsed GGUF tensor count to 36B in its automatic badge. This release is the Qwen3.6 35B-A3B MoE family: about 35B-class total parameters with roughly 3B active parameters per token.

Field	Value
model size	35B-A3B MoE
total parameters	35B class
active parameters	~3B class
architecture	qwen35moe
direct upstream GGUF	GestaltLabs/Qwen3.6-35B-A3B-NSC-ACE-SABER-GGUF-MTP
base family	Qwen/Qwen3.6-35B-A3B
local runtime format	ROCmFP4 Chadrock GGUF

Speed

Fresh HumanEval generation speed on AMD Ryzen AI Max+ 395 / Strix Halo with the included Ace Saber ROCmFP4 GGUF. This row uses the Chadrock Vulkan d2 quality profile at 32k context, f16 main/draft KV, draft-MTP depth 2, and deterministic decoding.

Metric	Result
HumanEval tasks	164
completion tokens generated	48,824
aggregate llama-server prompt speed	~543.56 tok/s
aggregate llama-server eval speed	~101.31 tok/s
peak live prompt speed	~544.93 tok/s
peak live eval speed	~104.35 tok/s
run context	32,768

These numbers come from the 2026-06-07 EvalPlus HumanEval rerun that produced the 155/164 base pass@1 and 148/164 HumanEval+ pass@1 result. The rerun replaced the older ROCm/q8/q4/depth-3 serving recipe because local tuning found the faster and more stable Chadrock profile on this Strix Halo box is Vulkan target+draft, f16 KV, larger batch/ubatch, and draft-MTP depth 2.

HumanEval

This model also posts an exceptional HumanEval result for a local GGUF run:

Model / row	HumanEval base pass@1	HumanEval+ pass@1
Chadrock-35B Ace Saber ROCmFP4, 32k Vulkan d2 rerun	155/164 = 94.51%	148/164 = 90.24%
earlier Chadrock-35B Ace Saber ROCmFP4 run	157/164 = 95.73%	149/164 = 90.85%
recorded stock Qwen3.6-27B UD-Q8_K_XL	154/164 = 93.90%	149/164 = 90.85%

The fresh 32k rerun still beats the recorded stock 27B row on base HumanEval, while the older row remains one task higher on HumanEval+.

BigCodeBench-Hard

The same tuned Chadrock Vulkan d2 family was also run on BigCodeBench-Hard-Instruct:

Benchmark	Result
BigCodeBench-Hard-Instruct pass@1	47/148 = 31.76%
generation wall time	799 s
aggregate prompt speed	~624.06 tok/s
aggregate generation speed	~100.12 tok/s

This is a harder instruction-coding benchmark than HumanEval and is included as a sanity check that the speed-tuned runtime still produces usable code under a broader task mix.

Run With llama-server

Build Charlie's custom llama.cpp once, download this GGUF, then run:

HSA_OVERRIDE_GFX_VERSION=11.5.1 \
GGML_HIP_ENABLE_UNIFIED_MEMORY=1 \
/path/to/rocmfp4-llama/build-strix-rocmfp4/bin/llama-server \
  -m Qwen3.6-35B-A3B-NSC-ACE-SABER-MTP-F16-to-ROCmFP4-STRIX_LEAN.gguf \
  --alias chadrock-35b-ace-saber \
  --host 127.0.0.1 \
  --port 8080 \
  --jinja \
  -c 262144 \
  --reasoning off \
  --reasoning-format none \
  --reasoning-budget 0 \
  -ngl 999 \
  -fa on \
  -sm row \
  -dev Vulkan0 \
  -b 8192 \
  -ub 2048 \
  -t 16 \
  -tb 32 \
  -ctk f16 \
  -ctv f16 \
  --spec-type draft-mtp \
  --spec-draft-device Vulkan0 \
  --spec-draft-ngl all \
  --spec-draft-type-k f16 \
  --spec-draft-type-v f16 \
  --spec-draft-threads 16 \
  --spec-draft-threads-batch 32 \
  --spec-draft-n-max 2 \
  --spec-draft-n-min 0 \
  --spec-draft-p-min 0.0 \
  --spec-draft-p-split 0.10 \
  --poll 100 \
  --poll-batch 1 \
  --spec-draft-poll 1 \
  --spec-draft-poll-batch 1 \
  --temp 0 \
  --min-p 0.0 \
  --top-p 0.9 \
  --top-k 20 \
  --repeat-penalty 1.0 \
  --seed 123 \
  --cache-ram 0 \
  --parallel 1 \
  --no-mmproj \
  --metrics

Use --parallel 1 for MTP. Multi-slot serving changes the MTP behavior and is not the intended profile. The benchmark table above used -c 32768 for the EvalPlus rerun. For general local coding and agent work, use the full -c 262144 context unless your hardware cannot hold it.

About Ace Saber

The source checkpoint is GestaltLabs/Qwen3.6-35B-A3B-NSC-ACE-SABER, based on Qwen/Qwen3.6-35B-A3B. The direct GGUF-MTP source is GestaltLabs/Qwen3.6-35B-A3B-NSC-ACE-SABER-GGUF-MTP.

Training order:

Qwen3.6-35B-A3B -> NSC-ACE -> SABER

NSC-ACE uses multiple steered rollouts from the same model and rewards convergence across latent behavior modes, especially for tool-call structure, reasoning wrappers, self-consistency, and avoiding repeated loops.

SABER is the final calibration pass. The source model card reports 98.33% HarmBench-300 compliance and final KLD 0.025383937664711.

About Chadrock / ROCmFP4

Charlie's ROCmFP4 method adds AMD-focused GGUF tensor formats and backend paths to llama.cpp.

ROCmFP4 is not stock Q4, MXFP4, or NVFP4. It uses custom Codebook10 4-bit weights, finite unsigned E4M3 scale semantics, tensor-aware presets, ROCm/HIP kernels, Vulkan shader support, and MTP regression guards.

Why it matters: Ryzen AI Max+ 395 / Strix Halo has a large unified-memory pool, but decode speed still depends heavily on bandwidth, tensor layout, and draft-token acceptance. ROCmFP4 is designed to make this class of AMD machine fast enough for serious local long-context use.

Build The Required llama.cpp

The GGUF is already provided. You only need to build the custom llama.cpp server once:

git clone https://github.com/charlie12345/rocmfp4-llama.git
cd rocmfp4-llama
git checkout mtp-rocmfp4-strix
env JOBS=16 scripts/build-strix-rocmfp4-mtp.sh

The server binary will be here:

build-strix-rocmfp4/bin/llama-server

File

File	SHA256
Qwen3.6-35B-A3B-NSC-ACE-SABER-MTP-F16-to-ROCmFP4-STRIX_LEAN.gguf	6a635d1d8ac4af8f2c4ca6ff528bc6bad9b3a6d45e8630ef6e5728f04898eeed

Credits

• @DJLougen: Ace Saber / NSC-ACE SABER model build.
• charlie12345 / @Italianclownz: ROCmFP4 llama.cpp fork, Strix Halo build path, and AMD-focused MTP runtime work.
• Qwen: base Qwen3.6-35B-A3B model.

Notes

This is an experimental AMD ROCmFP4/MTP build. Performance depends on driver version, clocks, prompt shape, MTP acceptance, and serving flags. The numbers above are local reproducible measurements, not universal llama.cpp claims.