Chadrock 35b Ace Saber Rocmfp4 Mtp

Chadrock 35B Ace Saber

New Chadrock v2 llama.cpp and config!

Chadrock-35B Ace Saber ROCmFP4 MTP + ROCmFPX MoEQuality

Chadrock-35B Ace Saber is a local AMD Strix Halo release with two additive serving lanes: the original ROCmFP4/MTP speed lane, and a new ROCmFPX MoEQuality-7.07 BPW lane for quality-sensitive coding and agent work. The original lane remains here unchanged; the FPX file is added as a second downloadable GGUF.

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 published measurements below use the Chadrock v2 ROCmFPX llama.cpp runner family from ciru-ai/ROCmFPX, with the exact ROCmFP4 and ROCmFPX runner references listed in the settings section.

These GGUFs will not run correctly with stock llama.cpp. Use the Chadrock v2 ROCmFPX runner because these files use ROCmFP4/ROCmFPX tensor types and MTP serving controls that upstream llama.cpp does not currently understand.

The model files are 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
• ROCmFPX MoEQuality-7.07 BPW for the additive ultra-quality lane
• 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 speed GGUF plus ROCmFPX MoEQuality GGUF, with 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.

Release Lanes

Lane	File	Best use	Notes
ROCmFP4 speed / vision lane	Qwen3.6-35B-A3B-NSC-ACE-SABER-MTP-F16-to-ROCmFP4-STRIX_LEAN.gguf	Fast local ACE/SABER serving with projector support	Original published lane; kept in place.
ROCmFPX MoEQuality lane	CHADROCK-35B-Ace-Saber-MTP-ROCmFPX-MoEQuality-7.07BPW.gguf	Higher-quality coding and agentic work	New additive quality lane, built from the H33 ACE/SABER MoEQuality recipe.
Vision projector	mmproj-CHADROCK-35B-Ace-Saber-F32.mmproj	Image-text prompts	Same projector can be used with either language GGUF when memory allows.

ROCmFPX MoEQuality-7.07 BPW Lane

The new FPX lane was built as a quality-focused MoE recipe for the same ACE/SABER 35B-A3B MTP model. The full quantization log reports the 753-tensor qwen35moe topology at 29945.07 MiB (7.07 BPW).

Best local task-eval profile for this lane is Vulkan, one slot, prompt cache disabled, deterministic sampling, MTP on, f16 draft KV, and the current draft cap profile described below.

Quality Metrics

Benchmark / row	Setting	Result	Runtime / speed
HermesAgent-20 official	ROCmFPX MoEQuality, Vulkan, 64k, q8_0 target KV, f16 draft KV, draft-MTP n3	0.82 avg score	506.60 s
HermesAgent-20 category scores	memory / workspace / skills / scheduling / delegation	1.00 / 0.80 / 0.80 / 0.75 / 0.73	same run
HumanEval / EvalPlus	ROCmFPX MoEQuality cap2 256k profile, deterministic	155/164 = 94.51% base, 148/164 = 90.24% plus	672.05 s, 71.31 tok/s generation

Useful local comparisons:

Comparison	Result
Stock Unsloth Qwen3.6 35B Q6 K XL, vanilla Vulkan, HermesAgent-20 draftKV16	0.71 avg score, so the FPX MoEQuality lane is +0.11 absolute on this agent bench.
Original ROCmFP4 ACE/SABER 32k HumanEval guard	155/164 base and 148/164 plus, matching the new FPX HumanEval row.
Stock Unsloth Qwen3.6 35B Q6 K XL HumanEval reference	157/164 base and 150/164 plus; FPX is within two tasks while generating slightly faster in the recorded EvalPlus row (71.31 vs 68.84 tok/s).

Speed Metrics

The tuned Vulkan cap2 profile uses f16 target KV, f16 draft KV, speculative.n_max=2, p_min=0.25, p_split=0.10, gen512, and prompt cache disabled:

Context	PP tok/s	TG tok/s	TTFP	Total time	Draft accepted
2k	819.99	71.37	2.84 s	10.02 s	290/412 = 70.4%
4k	770.17	69.44	5.33 s	12.71 s	292/416 = 70.2%
8k	771.23	69.55	10.64 s	18.00 s	294/415 = 70.8%
16k	737.70	69.21	22.22 s	29.62 s	302/405 = 74.6%

Matched 8k/gen512 comparison against stock Unsloth Qwen3.6 35B Q6 K XL, same prompt, Vulkan, f16 target/draft KV, MTP on, n_max=2, and deterministic sampling:

Row	PP tok/s	TG tok/s	TTFP	Total time	Draft accepted
ROCmFPX MoEQuality 7.07 BPW	771.23	69.55	10.64 s	18.00 s	294/415 = 70.8%
Stock Unsloth Q6 K XL	825.35	64.39	9.93 s	17.89 s	287/434 = 66.1%

On that matched row, the FPX quality lane is 8.0% faster on decode (69.55 vs 64.39 tok/s), while Q6 keeps a prefill and TTFP edge and the short 512-token wall time is nearly tied.

The first backend sweep for the FPX lane used cap4, n_max=4, and f16 target/draft KV. Vulkan was the better serving path for decode and draft acceptance:

Context	Vulkan TG	ROCm TG	Vulkan accept	ROCm accept
2k	95.82	33.25	402/428 = 93.9%	214/987 = 21.7%
4k	70.90	34.58	358/567 = 63.1%	246/872 = 28.2%
8k	56.87	41.45	325/670 = 48.5%	287/765 = 37.5%
16k	59.66	37.86	342/644 = 53.1%	278/820 = 33.9%

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

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

Runner references used for the published rows:

ROCmFP4 speed lane: ciru-ai/ROCmFPX commit 7aa484a2f0a504dc612a3d74a068024f3e6d6353
ROCmFPX MoEQuality lane: ciru-ai/ROCmFPX branch ciru-ai/fp6-vulkan-dequant-speed, commit fd6555ffb36dcef6419757189012c8c4fd31f91f

Recommended build for current serving: ciru-ai/ROCmFPX main commit 647fd52965a401d7fa8e035fc76a16c94216a794.

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.

For...