Qwable 5 27B Chadrock v2 ROCmFP4 is a Chadrock ROCmFP4/MTP GGUF release of DJLougen/Qwable-5-27B-Coder, tuned for AMD Ryzen AI Max+ 395 / Strix Halo systems.

Qwable is an agentic coding tune of Qwen3.6 27B built for repository work, terminal feedback, tool-use style prompts, and long coding turns. This release keeps that Qwable behavior and converts it into a compact Chadrock ROCmFP4 GGUF for fast served MTP on AMD unified-memory hardware.

This GGUF does not run correctly with stock upstream llama.cpp. It requires the Chadrock ROCmFPX / llama.cpp runner described here:

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

The setup page includes the pinned build commands, runner tag, launch config, request-level speculative controls, and validation checks needed to reproduce the speed rows below.

What This Build Is For

Local coding-agent and repository-work prompts on Strix Halo.
Fast one-slot served MTP with request-level speculative.n_max control.
Text-only llama-server use with a compact 14 GB ROCmFP4 GGUF.
Benchmarking Qwable against vanilla Qwen3.6 27B Q4 MTP under the same Chadrock runtime.

Lineage

text

unsloth/Qwen3.6-27B
  -> DJLougen/Qwable-5-27B-Coder
       training focus:
         - Claude Fable 5 coder-agent traces
         - Kimi 2.7 Coder traces
         - repository work, terminal workflows, tool-use style coding
  -> Qwable 5 27B Chadrock v2 ROCmFP4

The source model card for Qwable is Apache-2.0 and describes the model as a Qwen3.6-based coder-agent tune for real coding loops: inspect, edit, verify, and recover.

File

File	Size	SHA256
`Qwable-5-27B-Chadrock-v2-ROCmFP4.gguf`	`14,817,252,416 bytes`	`87818044a04cdfc0b5b7af897f0cf707273b7e4175b52a55cde8bd5c4bc12e3d`

Best Known Strix Halo Settings

Use the pinned ROCmFPX setup page for the full build and launch guide:

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

Current tested text-speed profile:

text

backend: Vulkan0 target + Vulkan0 draft
context: 131072 tested locally, 262144 model context metadata
batch / ubatch: 2048 / 512
target KV: q8_0 / q8_0
draft KV: f16 / f16
MTP: draft-mtp
startup draft cap: n_max=6, n_min=0, p_min=0.0, p_split=0.20
short-context request policy: speculative.n_max=6
mid-context request policy: speculative.n_max=4
long-context request policy: speculative.n_max=2
serving: one slot, metrics on, text-only for speed runs
sampler: temperature=0, top_p=0.95, top_k=20

Start the server with a draft cap at least as high as the deepest policy you plan to test. A request can lower speculative.n_max, but it cannot raise it above the server startup cap.

Run With Chadrock ROCmFPX

Build the runner from the pinned runner commit:

bash

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

Launch the text-speed profile:

bash

MODEL=/path/to/Qwable-5-27B-Chadrock-v2-ROCmFP4.gguf \
ALIAS=qwable-5-27b-chadrock-v2-rocmfp4 \
PORT=18180 \
CTX_SIZE=131072 \
DEVICE=Vulkan0 \
SPEC_DRAFT_DEVICE=Vulkan0 \
BATCH_SIZE=2048 \
UBATCH_SIZE=512 \
CACHE_TYPE_K=q8_0 \
CACHE_TYPE_V=q8_0 \
CACHE_TYPE_K_DRAFT=f16 \
CACHE_TYPE_V_DRAFT=f16 \
SPEC_DRAFT_N_MAX=6 \
SPEC_DRAFT_N_MIN=0 \
SPEC_DRAFT_P_MIN=0.0 \
SPEC_DRAFT_P_SPLIT=0.20 \
CPENT=-1 \
CTXCP=0 \
CACHE_RAM=8192 \
STRICT_BENCH=0 \
scripts/run-rocmfpx-mtp-server.sh

Example /completion request:

bash

curl -sS http://127.0.0.1:18180/completion \
  -H 'Content-Type: application/json' \
  -d '{
    "prompt": "Write a concise technical note about ROCmFPX MTP serving.",
    "n_predict": 512,
    "temperature": 0,
    "ignore_eos": true,
    "speculative.n_max": 6,
    "speculative.n_min": 0,
    "speculative.p_min": 0.0
  }'

Use --parallel 1 / one slot for MTP speed testing. Multi-slot serving changes draft-MTP behavior and is not the intended profile for these numbers.

Speed Tests vs Vanilla Q4

All numbers below were measured locally on AMD Ryzen AI Max+ 395 / Strix Halo with the same Chadrock Vulkan llama-server runtime, same prompt files, same API harness, same q8_0 target KV, same f16 draft KV, same b2048/u512, temperature=0, 512 generated tokens, and one-slot draft-MTP serving.

The vanilla baseline is the non-Chadrock model file /srv/ssd/sn850x/models/froggeric-Qwen3.6-27B-MTP-GGUF/Qwen3.6-27B-Q4_K_M-mtp.gguf, served on the same Chadrock runtime. These rows compare model/quant behavior under one runtime; they are not stock upstream llama.cpp runtime rows.

Model	Prompt tokens	Request draft setting	Decode tok/s	Prefill tok/s	Full request time	Draft accepted
Qwable 5 27B Chadrock v2 ROCmFP4	3,946	`n6`	`53.25`	`305.33`	`22.55 s`	`438 / 438`
Qwable 5 27B Chadrock v2 ROCmFP4	3,946	`n4`	`45.66`	`306.15`	`24.11 s`	`408 / 408`
Vanilla Qwen3.6 27B Q4_K_M MTP	3,946	`n4`	`27.82`	`264.04`	`33.36 s`	`363 / 588`
Vanilla Qwen3.6 27B Q4_K_M MTP	3,946	`n6`	`25.61`	`263.88`	`34.95 s`	`382 / 771`
Qwable 5 27B Chadrock v2 ROCmFP4	18,766	`n4`	`31.63`	`276.42`	`84.10 s`	`377 / 532`
Vanilla Qwen3.6 27B Q4_K_M MTP	18,766	`n4`	`34.45`	`241.13`	`92.71 s`	`402 / 432`

Readout

At the short 3,946-token prompt point, Qwable 5 Chadrock v2 is much faster even when draft depth is controlled. With both models at n4, Qwable reaches 45.66 tok/s vs vanilla 27.82 tok/s, a 64.1% decode speedup.

The best short-context Qwable row is 53.25 tok/s at n6 because it accepted every generated draft token on that prompt. Giving vanilla the same n6 setting made it slower, not faster: vanilla generated 771 draft tokens but accepted only 382.

At the 18,766-token prompt point, vanilla Q4 decodes faster under n4, but Qwable still prefills faster and finishes the full 512-token request sooner overall: 84.10 s vs 92.71 s.

Benchmark Provenance

Local run directory:

/home/crown/bench-results/llama/qwable-5-27b-chadrock-v2-rocmfp4-vs-vanilla-q4-20260621/

Imported benchmark labels:

text

qwable-5-27b-chadrock-v2-rocmfp4-dynamic-n6-ctx4k-g512
qwable-5-27b-chadrock-v2-rocmfp4-n4-ctx4k-g512
qwable-5-27b-chadrock-v2-rocmfp4-dynamic-n4-ctx16k-g512
vanilla-qwen3-6-27b-q4-k-m-mtp-n4-ctx4k-g512
vanilla-qwen3-6-27b-q4-k-m-mtp-n6-ctx4k-g512
vanilla-qwen3-6-27b-q4-k-m-mtp-n4-ctx16k-g512

The local benchmark store verified cleanly after import.

Credits

DJLougen: Qwable 5 27B Coder source model and coder-agent training.
Unsloth and Qwen: Qwen3.6 27B base model path used by the source checkpoint.
froggeric: Qwen3.6 27B MTP GGUF vanilla Q4 baseline used for local comparison rows.
Ciru / Chadrock ROCmFPX: ROCmFP4 GGUF conversion, Strix Halo serving profile, and pinned runner setup.

Notes

This is an experimental AMD ROCmFP4/MTP release for local evaluation and runtime experimentation. Speeds are hardware-sensitive and depend on driver version, clocks, prompt shape, KV cache settings, and draft-token acceptance.

jcbtc/qwable-5-27b-chadrock-v2-rocmfp4

Qwable 5 27B Chadrock v2 ROCmFP4

What This Build Is For

Lineage

File

Best Known Strix Halo Settings

Run With Chadrock ROCmFPX

Speed Tests vs Vanilla Q4

Readout

Benchmark Provenance

Credits

Notes

jcbtc/
qwable-5-27b-chadrock-v2-rocmfp4