Local AI Registry

Recipe

deepseek-v4-flash-0731-mxfp4-radeon-ai-pro-r9700-32gb-llama-cpp-tp1

deepseek-v4-flash-0731-mxfp4-radeon-ai-pro-r9700-32gb-llama-cpp-tp1
JSON APIFind compatible recipes

Observed LocalMaxxing leaderboard run. Evidence for compatibility, not an executable launch contract.

Record

Status
candidate
Source
localmaxxing
Engine
llama.cpp
Engine version
ROCmFPX fork (charlie12345/ROCmFPX) gfx1201 build
Accelerators
1
Tensor parallel
1
Context tokens
32,768
Max concurrency
1
chat
unknown
reasoning
unknown
tools
unknown
vision
unknown

Hugging Face model card

Identity

https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731
Repository
deepseek-ai/DeepSeek-V4-Flash-0731
Status
known
Link type
Exact Hub repository

Public Hugging Face repository confirmed by the Hub API.

Candidate: useful compatibility or speed evidence. The registry does not offer Run until promotion requirements are met.

Observed configuration

llama.cpp

Evidence only · candidate · reference

Candidate evidence — not a Run contract

Source
https://www.localmaxxing.com/en/runs/cmt4wezih03agpu01797e8l9m

Observed source tokens

Mechanical split of the source command. Unverified against the engine CLI. The registry does not offer Run for this recipe.

  1. llama-server
  2. (ROCmFPX
  3. gfx1201)
  4. --model
  5. DeepSeek-V4-Flash-ROCMFP2-STRIX.gguf
  6. --model-draft
  7. DeepSeek-V4-Flash-DSpark-draft-Q4RMFP4-denseF16.gguf
  8. -ngl
  9. 99
  10. -c
  11. 32768
  12. --draft-p-min
  13. 0.60
  14. (R9700=hip:0,
  15. Strix
  16. Halo
  17. iGPU=hip:1,
  18. peer
  19. access
  20. enabled,
  21. DSpark
  22. q=4,
  23. top-4
  24. experts/token)
  25. --port
  26. 18085
FlagValue
--modelDeepSeek-V4-Flash-ROCMFP2-STRIX.gguf
--model-draftDeepSeek-V4-Flash-DSpark-draft-Q4RMFP4-denseF16.gguf
-ngl99
-c32768
--draft-p-min0.60
--port18085

Measured speed

ConcurrencyContextPrefillDecodeTTFT msStatusSweep
132,768—52—observeddeepseek-v4-flash-0731-mxfp4-radeon-ai-pro-r9700-32gb-llama-cpp-tp1-sweep

Remaining fields

Identity, launch, related records, and measured speed are shown above. This is the rest of the normalized record.

hardware count
1
hardware id
radeon-ai-pro-r9700-32gb
id
deepseek-v4-flash-0731-mxfp4-radeon-ai-pro-r9700-32gb-llama-cpp-tp1
model instance id
deepseek-ai-deepseek-v4-flash-0731--mxfp4
recipe source
localmaxxing
schema version
local-ai-registry/v1
speed sweep ids
deepseek-v4-flash-0731-mxfp4-radeon-ai-pro-r9700-32gb-llama-cpp-tp1-sweep
status
candidate

capabilities

engine

name
llama.cpp
version
ROCmFPX fork (charlie12345/ROCmFPX) gfx1201 build

serving

max concurrency
1
max context tokens
32,768
tensor parallel
1
Provenance & metadata (3)

facts

capabilities.chat · provenance · captured at
2026-08-30T09:10:02Z
capabilities.chat · reason capability-not-verified
capabilities.chat · state unknown
capabilities.reasoning · provenance · captured at
2026-08-30T09:10:02Z
capabilities.reasoning · reason capability-not-verified
capabilities.reasoning · state unknown
capabilities.tools · provenance · captured at
2026-08-30T09:10:02Z
capabilities.tools · reason capability-not-verified
capabilities.tools · state unknown
capabilities.vision · provenance · captured at
2026-08-30T09:10:02Z
capabilities.vision · reason capability-not-verified
capabilities.vision · state unknown
engine.graph mode · provenance · captured at
2026-08-30T09:10:02Z
engine.graph mode · reason runtime-detail-not-published
engine.graph mode · state unknown
serving.kv cache tokens · provenance · captured at
2026-08-30T09:10:02Z
serving.kv cache tokens · reason kv-cache-capacity-not-published
serving.kv cache tokens · state unknown

metadata

localmaxxing · hardware label
Radeon AI Pro R9700
localmaxxing · notes
DeepSeek V4 Flash ROCmFP2, HYBRID dual-device: AMD Radeon AI PRO R9700 (gfx1201, 32GB) = hip:0 as primary compute device, PLUS the Ryzen AI Max 395 Strix Halo integrated GPU (Radeon 8060S, gfx1151) = hip:1 sharing 128GB unified system memory, connected over Oculink with GPU peer access enabled. The hardware schema is a discriminated union and cannot express discrete+unified together, so only the R9700 is listed in the gpus array; the iGPU's participation is disclosed here. Weights DeepSeek-V4-Flash-ROCMFP2-STRIX.gguf (102.3 GB, engine-native ROCmFP2 format; no matching LocalMaxxing quant enum, MXFP4 chosen as nearest FP4 label) + DSpark speculative draft DeepSeek-V4-Flash-DSpark-draft-Q4RMFP4-denseF16.gguf (11.3 GB, q=4). Decode-only measurement: median of 3 timed 128-token generations after 2 warmups, temperature 0, ~1956-token prompt, single concurrency, server-authoritative timing: 51.7 / 52.0 / 52.0 tok/s (median 52.0). DSpark acceptance ~0.99, outputs byte-identical across runs. IMPORTANT APPROXIMATION: this run used top-4 experts/token to reproduce a published headline figure; the quality-default 6-experts/token profile measured ~45.5 tok/s on the same rig. Prefill throughput, TTFT, power draw and peak VRAM were not recorded for this run, and the source blog's 415 tok/s prefill was NOT reproduced (decode only). Cold-prefill enforcement was not instrumented.
localmaxxing · observed command
llama-server (ROCmFPX gfx1201) --model DeepSeek-V4-Flash-ROCMFP2-STRIX.gguf --model-draft DeepSeek-V4-Flash-DSpark-draft-Q4RMFP4-denseF16.gguf -ngl 99 -c 32768 --draft-p-min 0.60 (R9700=hip:0, Strix Halo iGPU=hip:1, peer access enabled, DSpark q=4, top-4 experts/token) --port 18085
localmaxxing · run id
cmt4wezih03agpu01797e8l9m
localmaxxing · tokenized · arguments
llama-server, (ROCmFPX, gfx1201), --model, DeepSeek-V4-Flash-ROCMFP2-STRIX.gguf, --model-draft, DeepSeek-V4-Flash-DSpark-draft-Q4RMFP4-denseF16.gguf, -ngl, 99, -c, 32768, --draft-p-min, 0.60, (R9700=hip:0,, Strix, Halo, iGPU=hip:1,, peer, access, enabled,, DSpark, q=4,, top-4, experts/token), --port, 18085
localmaxxing · tokenized · fidelity
faithful

provenance

captured at
2026-08-30T09:10:02Z

sources

captured atkindurl
2026-08-30T09:10:02Znormalized-recipewww.localmaxxing.com/en/runs/cmt4wezih03agpu01797e8l9m ↗