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

New Chadrock v2 llama.cpp and config!

Chadrock-35B Ace Saber ROCmFP4 MTP

Chadrock-35B Ace Saber is a ROCmFP4/MTP GGUF for AMD Ryzen AI Max+ 395 / Strix Halo systems. This release also includes a tested Qwen3.6 vision projector, so the same full Chadrock language GGUF can run image-text prompts when launched with --mmproj.

The model behavior comes from the Ace Saber build by @DJLougen. The current speed numbers use the pinned Chadrock v2 ROCmFPX llama.cpp build from ciru-ai/ROCmFPX, with the request-level MTP controls described below.

This GGUF will not run correctly with stock llama.cpp. Use the pinned Chadrock v2 ROCmFPX runner because this file uses ROCmFP4 tensor types and MTP serving controls 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 plus separate GGUF-format vision projector

Vision Support

Vision is provided by mmproj-CHADROCK-35B-Ace-Saber-F32.mmproj, a GGUF-format Qwen3VL projector converted from the restored Qwen3.6 visual tower sidecar in the upstream Ace Saber release.

This does not replace the language model and does not disable MTP. The validated command used the full Chadrock ROCmFP4 GGUF with native --spec-type draft-mtp enabled and added the projector with --mmproj.

Local validation used two generated images whose answers were not present in the prompt:

Image gate	Expected	Result
gate_a.png	CIRU-742, red square, blue circle	passed
gate_b.png	HALO-319, orange triangle, purple star	passed

The same gate fails without a projector, so this is a real image-read check rather than a metadata-only claim.

Chadrock v2 Speed

Best current text-only profile on AMD Ryzen AI Max+ 395 / Strix Halo: pinned Chadrock v2 ROCmFPX llama.cpp, Vulkan0 target plus Vulkan0 draft, f16/f16 target and draft KV, one slot, prompt cache disabled, no multimodal projector, deterministic decoding, and request policy speculative.n_max=4, speculative.n_min=0, speculative.p_min=0.25.

Measurement	Prompt tokens	Generated tokens	Decode tok/s	Prefill tok/s	Total time	Draft accepted
Chadrock v2 best, gen512	3,946	512	143.08	1072.34	7.26 s	408 / 408
Chadrock v2 repeat, gen2048	3,946	2048	141.77	1064.16	18.16 s	1637 / 1637
Same-run no-draft control, gen512	3,946	512	72.57	1064.49	10.77 s	0 / 0
Same-run no-draft control, gen2048	3,946	2048	72.04	1067.18	32.13 s	0 / 0

Against the same-run no-draft control, the new Chadrock v2 MTP config is 1.97x faster in decode at both gen512 and gen2048: 143.08 vs 72.57 tok/s, and 141.77 vs 72.04 tok/s.

Compared with the older Chadrock v1 card-speed profile (~101.31 tok/s aggregate HumanEval eval speed), the new best served text row is about 1.41x faster (+41.2%). Compared with the older uncached Chadrock TG64 served row (78.28 tok/s server eval), the new gen2048 row is about 1.81x faster (+81.1%). Those older rows are not identical prompts, so treat them as release-to-release runner evidence rather than a strict apples-to-apples benchmark pair.

The older 2026-06-07 HumanEval rerun is still useful as a quality guard: it generated 48,824 completion tokens across 164 tasks at ~101.31 tok/s aggregate llama-server eval speed and produced the 155/164 base pass@1 and 148/164 HumanEval+ pass@1 result below.

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.

Best Settings / Advanced Setup

For the pinned runner build, copy-paste build commands, request-level speculative controls, and the 35B/27B reproduction notes, use the advanced Ciru setup page:

https://llm.ciru.ai/chadrock-rocmfpx/

The current pinned runner build is:

ciru-ai/ROCmFPX commit: 7aa484a2f0a504dc612a3d74a068024f3e6d6353
historical score tag: chadrock-rocmfp4-mtp-scores-20260621

For the fastest measured text-only ACE/SABER path on Strix Halo, use:

backend: Vulkan0 target + Vulkan0 draft
context: 32768
batch / ubatch: 2048 / 512
target KV: f16 / f16
draft KV: f16 / f16
MTP: draft-mtp, n_max=4, n_min=0, p_min=0.25, p_split=0.10
serving: one slot, prompt cache disabled for benchmarks, --no-mmproj for text speed
sampler: temperature=0, top_p=0.95, top_k=20

That is the profile that produced 143.08 tok/s at gen512 and repeated at 141.77 tok/s at gen2048 on the 3946-token text prompt. For image-text use, remove --no-mmproj and add --mmproj mmproj-CHADROCK-35B-Ace-Saber-F32.mmproj; the headline speed row above is text-only.

Run With llama-server

Build the pinned Chadrock v2 ROCmFPX llama.cpp once, download this GGUF, then run this text-speed profile from the ROCmFPX checkout:

./build-strix-rocmfp4/bin/llama-server \
  -m /path/to/Qwen3.6-35B-A3B-NSC-ACE-SABER-MTP-F16-to-ROCmFP4-STRIX_LEAN.gguf \
  --alias chadrock-35b-ace-saber-rocmfp4-cap4 \
  --host 127.0.0.1 \
  --port 18180 \
  --jinja \
  -c 32768 \
  --reasoning off \
  --reasoning-format none \
  --reasoning-budget -1 \
  --no-context-shift \
  -dev Vulkan0 \
  -ngl 999 \
  -fa on \
  -b 2048 \
  -ub 512 \
  -t 16 \
  -tb 32 \
  -ctk f16 \
  -ctv f16 \
  --temp 0 \
  --top-p 0.95 \
  --top-k 20 \
  --seed 123 \
  --parallel 1 \
  --no-mmproj \
  --metrics \
  --no-webui \
  --cache-ram 8192 \
  --ctx-checkpoints 0 \

  --checkpoint-every-n-tokens -1 \

  --spec-type draft-mtp \
  --spec-draft-device Vulkan0 \
  --spec-draft-ngl all \
  --spec-draft-threads 16 \
  --spec-draft-threads-batch 32 \
  --spec-draft-type-k f16 \
  --spec-draft-type-v f16 \
  --spec-draft-n-max 4 \
  --spec-draft-n-min 0 \
  --spec-draft-p-min 0.25 \
  --spec-draft-p-split 0.10 \
  --no-spec-draft-backend-sampling \
  --spec-draft-poll 1 \
  --spec-draft-poll-batch 1

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 and --no-mmproj for the fastest text row. For image-text use, remove --no-mmproj and add the projector:

--mmproj /path/to/mmproj-CHADROCK-35B-Ace-Saber-F32.mmproj

For general local coding and agent work, you can raise context after validating memory headroom on your machine. The advanced page below keeps the copy-paste build and launch blocks current:

https://llm.ciru.ai/chadrock-rocmfpx/

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

Chadrock v2 uses the pinned Ciru ROCmFPX llama.cpp branch carrying the ROCmFP4 tensor/runtime work and request-level MTP serving controls.

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/ciru-ai/ROCmFPX.git
cd ROCmFPX
git checkout 7aa484a2f0a504dc612a3d74a068024f3e6d6353
env JOBS=16 scripts/build-strix-rocmfp4-mtp.sh llama-server llama-bench

The server binary will be here:

build-strix-rocmfp4/bin/llama-server

Files

File	SHA256
Qwen3.6-35B-A3B-NSC-ACE-SABER-MTP-F16-to-ROCmFP4-STRIX_LEAN.gguf	6a635d1d8ac4af8f2c4ca6ff528bc6bad9b3a6d45e8630ef6e5728f04898eeed
mmproj-CHADROCK-35B-Ace-Saber-F32.mmproj	1365c90e3f35cad9c33e09e67ff377af083631c19718ec4c22d251a54c24c6a7

Credits

• @DJLougen: Ace Saber / NSC-ACE SABER model build.
• Ciru / ciru-ai/ROCmFPX: pinned Chadrock v2 ROCmFPX llama.cpp runner, build path, request-level MTP controls, and public reproduction page.
• charlie12345 / @Italianclownz: original ROCmFP4 direction and AMD-focused runtime work that informed the Chadrock path.
• 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.