kingjones777/UI-Mate-27B-ROCmFP4-STRIX-GGUF

UI-Mate-27B — ROCmFP4 for AMD Strix Halo (gfx1151)

A ROCmFP4 build of tencent/UI-Mate-27B, Tencent's GUI agent, for AMD Ryzen AI Max+ 395 (Radeon 8060S, gfx1151). UI-Mate is a screenshot-in, tool-call-out model: it reads a screen and emits the next action to take. This repository holds one model file, the BF16 vision projector that goes with it, and a DFlash2 draft model for speculative decoding.

Metric	Result
Quantization	ROCmFP4 (STRIX recipe), output head and token embedding at Q6_K
Model size	14.09 GiB (15,127,582,592 bytes)
Tested hardware	AMD Ryzen AI Max+ 395 (Strix Halo), 128 GB unified
GPU	Radeon 8060S, gfx1151
ROCm version	7.2.4
GUI-agent harness decode	30.3 tok/s with the drafter, 13.9 tok/s without
Decode at 1,564 / 13,382 / 26,819 prompt tokens	37.9 / 24.4 / 24.4 tok/s
Prefill at 1,564 / 13,382 / 26,819 prompt tokens	586.1 / 570.1 / 547.3 tok/s
Tool calls, GUI-agent harness	✅ 17 of 17 valid
Tool calls, required mode	✅ 7 of 7
Perplexity (held out, n_ctx 4096, 5 chunks)	5.9876 (bartowski's Q4_K_M: 6.0938)
Prompt caching	on — the conversation prefix is kept between turns

All numbers were measured on 2026-11-07 with Super ROC build fd85671, one request at a time, cold prompt cache, fixed prompts, and the sampling the model's own generation config asks for (temperature 1.0, top-p 0.95, top-k 20, thinking on).

Why this build?

• A working GUI agent at usable speed. The official GUI-agent requests run at 30.3 tok/s with the bundled drafter, and every one of the 17 produced a valid tool call.
• Speculative decoding is worth 2.18× here. On the GUI-agent harness the same model file decodes at 13.9 tok/s without the drafter and 30.3 tok/s with it, with 42 % of drafted tokens accepted.
• Prompt caching is on, so a multi-turn agent loop does not reprocess the whole conversation on every step.
• One file, one projector, one drafter — everything needed to serve the agent is in this repository.

Quick start

hf download kingjones777/UI-Mate-27B-ROCmFP4-STRIX-GGUF --local-dir ui-mate

superroc serve -i 127.0.0.1 -p 8080 -j 1 --max-request-bytes 67108864 llm \
  -m UI-Mate-27B-Q4_0_ROCMFP4_STRIX-imatrix-headQ6K.gguf \
  --mmproj mmproj-tencent_UI-Mate-27B-bf16.gguf \
  --context 65536 --think on -t 1.0 --top-p 0.95 --top-k 20 --presence-penalty 0 \
  --speculative dflash2 --dflash-model Qwen3.6-27B-DFlash2-JetBrains-Q4_K_M.gguf \
  --prompt-lookup

The --max-request-bytes 67108864 setting raises the request size limit to 64 MiB so that a screenshot payload fits. --prompt-lookup keeps the conversation prefix cached between turns.

Files

File	Size	Role
UI-Mate-27B-Q4_0_ROCMFP4_STRIX-imatrix-headQ6K.gguf	14.09 GiB (15,127,582,592 bytes)	model
mmproj-tencent_UI-Mate-27B-bf16.gguf	0.87 GiB (931,146,368 bytes)	vision projector
Qwen3.6-27B-DFlash2-JetBrains-Q4_K_M.gguf	1.06 GiB (1,143,006,848 bytes)	DFlash2 draft model

SHA-256:

File	sha256
UI-Mate-27B-Q4_0_ROCMFP4_STRIX-imatrix-headQ6K.gguf	09f9808d81c946b4b34187777bb78a4b0a591e10f63cf887b432f0e0baf8ba78
mmproj-tencent_UI-Mate-27B-bf16.gguf	19f276b345a8852200828c63d3b81656174bd6e60d23672b5e2a7ee668dc6a96
Qwen3.6-27B-DFlash2-JetBrains-Q4_K_M.gguf	68988bef15539ca8b16ffef6ff4ef61308ccdd443f961a245bf579cf2d5fac33

Measurements

Measured on AMD Ryzen AI Max+ 395 (Radeon 8060S, gfx1151), 128 GB, ROCm 7.2.4; one request at a time; cold prompt cache; fixed prompts.

Prompt	Prompt tokens	Prefill tok/s	Decode tok/s	Draft tokens accepted
1K	1,564	586.1	37.9	183 of 387
8K	13,382	570.1	24.4	146 of 422
16K	26,819	547.3	24.4	156 of 394

Code generation (six coding prompts, 256 tokens each): mean decode 27.3 tok/s, range 25.8 to 29.9 tok/s.

Tool calls in required mode (seven cases): 7 of 7 right.

GUI-agent harness

Tencent's 17 official GUI-agent requests, temperature 1.0, thinking on:

Configuration	Decode	Draft tokens accepted
Without the drafter	13.9 tok/s	—
With the DFlash2 drafter	30.3 tok/s	42 %

Tool calls: 17 of 17 valid.

Speculative decoding (DFlash2)

UI-Mate's release has no multi-token-prediction head, so there is no MTP draft file to load. Speculation here comes from a separate draft model: JetBrains' Qwen3.6-27B-DFlash2, an unofficial JetBrains adaptation of Inco AI's DFlash 2 for Qwen3.6-27B (Apache 2.0), converted to GGUF and quantized to Q4_K_M.

It works as a drafter for UI-Mate because UI-Mate keeps Qwen3.6-27B's vocabulary and architecture.

Acceptance on the fixed prompts above is 183 of 387 at 1K, 146 of 422 at 8K and 156 of 394 at 16K, and 42 % across the GUI-agent harness.

Architecture note

UI-Mate-27B is fine-tuned from Qwen3.6-27B and keeps the qwen35 architecture: 64 layers, of which 48 are Gated-DeltaNet and 16 are full attention. Only the 16 full-attention layers carry a KV cache, so context memory grows on a minority of the layers.

The model is dense — all 27B parameters are read on every token — so decode is close to bound by memory bandwidth. Speculative decoding clears that ceiling by producing several tokens per weight read, which is what the drafter buys here.

Sampling

The serve command uses the model's own generation config:

temperature	1.0
top_p	0.95
top_k	20
presence_penalty	0
thinking	on

Quantization methodology

The model file was requantized from bartowski's published Q8_0 GGUF of UI-Mate-27B with llama-quantize --allow-requantize, using the ROCmFP4 STRIX recipe and bartowski's published importance matrix. output.weight and token_embd.weight are at Q6_K.

The projector is bartowski's BF16 mmproj, unchanged.

The draft model is JetBrains' Qwen3.6-27B-DFlash2, converted to GGUF and quantized to Q4_K_M.

Not yet measured

Listed so nobody mistakes absence for a pass:

Test	Status
Prompts longer than 26,819 tokens	❓ not measured (the serve command allows 65,536)
Vision beyond the GUI-agent requests	❓ no separate projector accuracy test
Multi-step agentic loop (plan → act → observe → finish)	❓ not run
Independent reproduction	❓ none yet

Perplexity caveat. Both figures (this file 5.9876, bartowski's Q4_K_M 6.0938) come from the same held-out text, the same tool and the same settings (n_ctx 4096, 5 chunks). Perplexity measures next-token prediction on prose; it is not a proxy for agent behaviour, tool-call correctness, or screen-reading accuracy.

Known issues

1. These are ROCm files. ROCmFP4 is a runtime tensor format that exists only in the ROCmFPX fork of llama.cpp, so the model file runs on ROCm hardware only.
2. There is no MTP head. Drafting is done with the separate DFlash2 file, which must be passed to the server for speculative decoding to engage.

License and attribution

UI-Mate-27B and Qwen3.6-27B are published under Apache 2.0. The ROCmFP4 quantisation types are from the ROCmFPX fork of llama.cpp. This repository contains the quantised weights and the measurements above.

Acknowledgements

Tencent — UI-Mate-27B The base model and its fine-tuning are theirs, under Apache 2.0. This repository contributes quantisation and measurement only.

Qwen team — Qwen3.6-27B The model UI-Mate-27B is fine-tuned from. The vocabulary and architecture are also what make the DFlash2 draft model usable here.

bartowski — Q8_0 source, importance matrix, projector The published Q8_0 GGUF this file was requantized from, the published importance matrix used for the weighted quantisation, and the BF16 vision projector.

JetBrains and Inco AI — DFlash2 drafter DFlash 2 is Inco AI's work; the Qwen3.6-27B adaptation used as the draft model is JetBrains'. The GGUF conversion and Q4_K_M quantization of that draft model are this repository's.

Super ROC — the engine every number above was measured on.

If you use these files, please credit the authors above alongside this repository.