Local AI Registry

Recipe

qwen3-6-27b-q4-k-m-dgx-spark-gb10-128gb-llama-cpp-tp1

qwen3-6-27b-q4-k-m-dgx-spark-gb10-128gb-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
llama.cpp b8214 f7db3f378; Luce DFlash e534780
Accelerators
1
Tensor parallel
1
Context tokens
1,024
chat
unknown
reasoning
unknown
tools
unknown
vision
unknown

Hugging Face model card

Identity

https://huggingface.co/Qwen/Qwen3.6-27B
Repository
Qwen/Qwen3.6-27B
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/cmp7vw9de00goo401s6mzzoag

Observed source tokens

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

  1. /home/user/work/dflash-lucebox-gb10-spark3/venv/bin/python
  2. /home/user/work/dflash-lucebox-gb10-spark3/llama-benchy-fixed-serving-grid-20260514_164136/server_dflash-sherlock-thinkON.py
  3. --host
  4. 127.0.0.1
  5. --port
  6. 8180
  7. --target
  8. /home/user/work/dflash-lucebox-gb10-spark3/lucebox-hub/dflash/models/Qwen3.6-27B-Q4_K_M.gguf
  9. --draft
  10. /home/user/work/dflash-lucebox-gb10-spark3/lucebox-hub/dflash/models/draft
  11. --bin
  12. /home/user/work/dflash-lucebox-gb10-spark3/lucebox-hub/dflash/build/test_dflash
  13. --budget
  14. 18
  15. --ddtree-temp
  16. 1.05
  17. --max-ctx
  18. 1024
  19. --ctk
  20. f16
  21. --ctv
  22. f16
  23. --fa-window
  24. 2048
  25. --prefix-cache-slots
  26. 0
  27. --prefill-cache-slots
  28. 0
  29. --tokenizer
  30. Qwen/Qwen3.6-27B
  31. --verbose-daemon
  32. --ignore-eos-stop
  33. --ddtree-no-chain-seed
FlagValue
--host127.0.0.1
--port8180
--target/home/user/work/dflash-lucebox-gb10-spark3/lucebox-hub/dflash/models/Qwen3.6-27B-Q4_K_M.gguf
--draft/home/user/work/dflash-lucebox-gb10-spark3/lucebox-hub/dflash/models/draft
--bin/home/user/work/dflash-lucebox-gb10-spark3/lucebox-hub/dflash/build/test_dflash
--budget18
--ddtree-temp1.05
--max-ctx1024
--ctkf16
--ctvf16
--fa-window2048
--prefix-cache-slots0
--prefill-cache-slots0
--tokenizerQwen/Qwen3.6-27B

Environment

VariableValue
PYTHONPATH/home/user/work/dflash-lucebox-gb10-spark3/lucebox-hub/dflash/scripts

Measured speed

ConcurrencyContextPrefillDecodeTTFT msStatusSweep
11,024224.455.2570.4observedqwen3-6-27b-q4-k-m-dgx-spark-gb10-128gb-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
dgx-spark-gb10-128gb
id
qwen3-6-27b-q4-k-m-dgx-spark-gb10-128gb-llama-cpp-tp1
model instance id
qwen-qwen3-6-27b--q4-k-m
recipe source
localmaxxing
schema version
local-ai-registry/v1
speed sweep ids
qwen3-6-27b-q4-k-m-dgx-spark-gb10-128gb-llama-cpp-tp1-sweep
status
candidate

capabilities

engine

name
llama.cpp
version
llama.cpp b8214 f7db3f378; Luce DFlash e534780

serving

max context tokens
1,024
tensor parallel
1
Provenance & metadata (3)

facts

capabilities.chat · provenance · captured at
2026-08-30T09:26:09Z
capabilities.chat · reason capability-not-verified
capabilities.chat · state unknown
capabilities.reasoning · provenance · captured at
2026-08-30T09:26:09Z
capabilities.reasoning · reason capability-not-verified
capabilities.reasoning · state unknown
capabilities.tools · provenance · captured at
2026-08-30T09:26:09Z
capabilities.tools · reason capability-not-verified
capabilities.tools · state unknown
capabilities.vision · provenance · captured at
2026-08-30T09:26:09Z
capabilities.vision · reason capability-not-verified
capabilities.vision · state unknown
engine.graph mode · provenance · captured at
2026-08-30T09:26:09Z
engine.graph mode · reason runtime-detail-not-published
engine.graph mode · state unknown
serving.kv cache tokens · provenance · captured at
2026-08-30T09:26:09Z
serving.kv cache tokens · reason kv-cache-capacity-not-published
serving.kv cache tokens · state unknown
serving.max concurrency · provenance · captured at
2026-08-31T23:03:15Z
serving.max concurrency · reason server-capacity-not-evidenced
serving.max concurrency · state unknown

metadata

localmaxxing · backend
cuda
localmaxxing · batch size
1
localmaxxing · hardware label
GB10 Grace Blackwell
localmaxxing · notes
Qwen3.6-27B Q4_K_M GGUF on a single NVIDIA DGX Spark GB10. DFlash speculative decoding plus DDTree rollback/tree decoding (budget=18, ddtree-temp=1.05, no chain seed). LocalMaxxing row uses Sherlock/thinkON pp=128,tg=128,c=1, N=29 warm median: tokSOut 55.17 tok/s (mean 55.42, std 16.47, range 31.87-83.68); TTFT/e2e first-token median 570.38 ms; tokSPrefill 224.41 tok/s; tokSTotal 88.57 tok/s. Variability is prompt/mode dependent: across the four N=29 cells, AR median 10.11-10.69, +DF 16.77-33.77, +DDT 25.11-55.17 tok/s; DDTree adds 1.50-1.63x over +DF. This is the headline +DDT row; companion AR and +DF payloads are included for same-shape baseline context. Raw JSON, commands, token_id smokes, and three-tier AR/+DF/+DDT discussion: https://github.com/my-other-github-account/spark-bench-reproducers/tree/main/luce-ddtree/submission and https://github.com/my-other-github-account/spark-bench-reproducers/tree/main/luce-ddtree. By @banana_baeee.
localmaxxing · observed command
PYTHONPATH=/home/user/work/dflash-lucebox-gb10-spark3/lucebox-hub/dflash/scripts /home/user/work/dflash-lucebox-gb10-spark3/venv/bin/python /home/user/work/dflash-lucebox-gb10-spark3/llama-benchy-fixed-serving-grid-20260514_164136/server_dflash-sherlock-thinkON.py --host 127.0.0.1 --port 8180 --target /home/user/work/dflash-lucebox-gb10-spark3/lucebox-hub/dflash/models/Qwen3.6-27B-Q4_K_M.gguf --draft /home/user/work/dflash-lucebox-gb10-spark3/lucebox-hub/dflash/models/draft --bin /home/user/work/dflash-lucebox-gb10-spark3/lucebox-hub/dflash/build/test_dflash --budget 18 --ddtree-temp 1.05 --max-ctx 1024 --ctk f16 --ctv f16 --fa-window 2048 --prefix-cache-slots 0 --prefill-cache-slots 0 --tokenizer Qwen/Qwen3.6-27B --verbose-daemon --ignore-eos-stop --ddtree-no-chain-seed
localmaxxing · run id
cmp7vw9de00goo401s6mzzoag
localmaxxing · tokenized · environment · PYTHONPATH
/home/user/work/dflash-lucebox-gb10-spark3/lucebox-hub/dflash/scripts
localmaxxing · tokenized · fidelity
faithful

provenance

captured at
2026-08-30T09:26:09Z

sources

captured atkindurl
2026-08-30T09:26:09Znormalized-recipewww.localmaxxing.com/en/runs/cmp7vw9de00goo401s6mzzoag ↗