This repo contains the CHADROCK3.6 Coder releases for AMD Ryzen AI Max+ 395 / Strix Halo systems:
CHADROCK3.6-27B-Coder-MTP-ROCmFP4-STRIX_LEAN.gguf: the original dense 27B ROCmFP4/MTP Coder lane.
CHADROCK3.6-35B-A3B-Coder-MTP-ROCmFPX-MoEQuality-7.08BPW.gguf: the additive 35B A3B MoEQuality ROCmFPX/MTP Coder lane.
The existing 27B ROCmFP4 lane is preserved. The 35B MoEQuality file is added as a second download lane for users who want the larger A3B Coder behavior with a higher-quality ROCmFPX tensor mix.
CHADROCK3.6 Coder uses Qwopus3.6 Coder lineage, then converts the source models into Charlie's AMD-focused ROCmFP4 / ROCmFPX runtime formats. The public release names and artifact names are Chadrock names, while Qwopus stays explicit in lineage, base model metadata, and credits.
The 27B file is a compact 14 GB GGUF for local agentic coding, repository work, tool-use style prompts, and long-context experiments. The 35B A3B MoEQuality file is a larger 30 GB GGUF aimed at better tool-use/coder behavior while keeping draft-MTP serving available on unified-memory AMD hardware.
These GGUFs will not run correctly with stock llama.cpp. They need a pinned ciru-ai/ROCmFPX runner because the files use ROCmFP4 / ROCmFPX 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. Build the custom llama server once, download the files, and run the profile below.
Why This Build Exists
CHADROCK3.6 27B Coder is the Strix-focused Chadrock release of a dense agentic coding model lineage. It is intended for coding, tool use, debugging, structured developer workflows, and runtime experimentation on AMD hardware. Chadrock adds the AMD runtime piece:
ROCmFP4 Strix Lean tensor recipe
native draft-MTP serving
AMD ROCm/HIP backend path
262K context target
q4_0 KV cache profile for long local sessions
optional vision projector companion file
This release is best treated as a model/runtime pairing for Strix Halo rather than a generic GGUF quant.
In plain terms: Qwen provides the foundation models, Jackrong's Qwopus lines add Trace Inversion and coder/tool-use training, the upstream MTP GGUFs provide the MTP sources, and this release converts those lines into Strix-focused Chadrock runtime formats.
All numbers below were measured locally on AMD Ryzen AI Max+ 395 / Strix Halo.
Tool Eval Full 69 - 35B A3B ROCmFPX MoEQuality
The 35B A3B MoEQuality lane was run through the local Tool Eval full 69 deterministic tool-use suite with the raw llama.cpp endpoint, temperature=0, seed=42, parallel=1, --no-think, and --structured-response-format json_object.
Metric
Result
final score
72
points
100 / 138
scenarios
40 pass / 20 partial / 9 fail
median turn time
3708.1 ms
generated throughput during eval
21.14 tok/s
The run artifacts were audited for the earlier structured-output harness failure mode. No HTTP 400, sampler initialization, or backend schema-support failure signatures were found in the JSON, progress log, or generated report. The structured-output cases below are scored model behavior, not backend failures.
Category
Label
Score
A
Tool Selection
6 / 6 = 100%
B
Parameter Precision
6 / 6 = 100%
C
Multi-Step Chains
8 / 8 = 100%
D
Restraint & Refusal
5 / 6 = 83%
E
Error Recovery
5 / 6 = 83%
F
Localization
6 / 6 = 100%
G
Structured Reasoning
2 / 6 = 33%
H
Instruction Following
8 / 10 = 80%
I
Context & State
14 / 20 = 70%
J
Code Patterns
4 / 6 = 67%
K
Safety & Boundaries
18 / 26 = 69%
L
Toolset Scale
5 / 8 = 62%
M
Autonomous Planning
4 / 6 = 67%
N
Creative Composition
3 / 6 = 50%
O
Structured Output
6 / 12 = 50%
BigCodeBench Hard Instruct
Run
Result
bigcodebench-hard-instruct, calibrated
48/148 = 32.43% pass@1
The scored run used the local profile qwopus3.6-27b-coder-mtp-chadrock-rocmfp4-strix-lean in the June 13 full coding benchmark folder. That local profile name records source lineage and build path; the public release name is CHADROCK3.6 27B Coder.
Apples-to-Apples Q5_K_M Comparison
The cleanest decode-speed comparison is the same CHADROCK CLI guard run against the upstream Qwopus3.6 27B Coder MTP Q5_K_M GGUF and this Chadrock ROCmFP4 build, using the same prompts, runtime build, machine, and MTP guard harness.
Guard row
Upstream Q5_K_M decode
Chadrock ROCmFP4 decode
Decode uplift
short arithmetic prompt
17.8 tok/s
29.5 tok/s
1.66x
sustained regression-guard prompt
13.2 tok/s
22.6 tok/s
1.71x
Prompt processing is better represented by the no-cache long-context sweep below, where Chadrock ROCmFP4 measured 315.97 tok/s at 4K prompt tokens and 142.00 tok/s at 130K prompt tokens.
Long-Context Sweep
The no-cache forced context sweep generated 512 tokens at each context length:
Prompt tokens
Prompt speed
Decode speed
Draft accepted
4,131
315.97 tok/s
21.25 tok/s
314/779
8,227
308.66 tok/s
21.82 tok/s
329/728
16,419
286.62 tok/s
21.64 tok/s
344/666
32,803
251.76 tok/s
17.35 tok/s
335/701
65,571
201.49 tok/s
12.51 tok/s
329/726
130,467
142.00 tok/s
7.08 tok/s
305/823
These are local server measurements, not universal llama.cpp claims. Throughput depends heavily on driver version, clocks, prompt shape, KV cache settings, and MTP acceptance.
Best Settings / Advanced Setup
For the pinned runner build, copy-paste build commands, request-level speculative controls, and the latest Chadrock ROCmFP4 reproduction notes, use the advanced Ciru setup page:
Use the advanced page if you are testing the newer request-level ROCmFPX runner
or comparing against the separate Qwable 5 27B Coder ROCmFP4 speed lane. The
settings above are the best published-card settings for the actual GGUF in this
repo.
Run With llama-server
27B ROCmFP4 Lane
Build Charlie's custom llama.cpp once, download this GGUF and the projector file, then run:
Use --parallel 1 for this MTP profile. Multi-slot serving changes draft-MTP behavior and is not the intended configuration.
For text-only use, you may omit --mmproj.
For vision use, keep mmproj-F32.mmproj beside the main GGUF, but run with MTP off. In practice, that means using the vision projector and removing the --spec-* draft-MTP flags from the command.
The projector is a GGUF-format projector file with a .mmproj repo extension so Hugging Face's GGUF metadata badge tracks the 27B language model rather than the smaller projector.
35B A3B ROCmFPX MoEQuality Lane
For the 35B MoEQuality file, use a ROCmFPX runner with Vulkan support and the same Froggeric Qwen fixed chat template used for the local Tool Eval run:
For text-only serving, omit the projector. If you use the included mmproj-CHADROCK3.6-35B-A3B-Coder-MTP-F32.mmproj projector, validate your vision path separately and run with MTP off unless your local runner supports that combination.
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 4-bit tensor layouts, Codebook10 values, finite unsigned E4M3 scale semantics, tensor-aware Strix presets, ROCm/HIP kernels, Vulkan support, and MTP regression guards.
Why it matters: Strix Halo has a large unified-memory pool, but good local serving still depends on memory bandwidth, tensor layout, KV traffic, and draft-token acceptance. Chadrock is built for that exact hardware shape.
This is an experimental AMD ROCmFP4/MTP build. It is intended for local evaluation, coding workflows, and runtime experimentation on compatible AMD hardware.
