Local AI Registry

Recipe

qwen3-6-27b-q4-k-m-rtx-3090-24gb-llama-cpp-tp2-sl792yp8

qwen3-6-27b-q4-k-m-rtx-3090-24gb-llama-cpp-tp2-sl792yp8
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
Tom-TurboQuant-fork-b3401eaf9
Accelerators
2
Tensor parallel
2
Context tokens
65,536
Max concurrency
1
chat
unknown
reasoning
unknown
tools
unknown
vision
unknown

Hugging Face model card

Identity

https://huggingface.co/unsloth/Qwen3.6-27B-GGUF
Repository
unsloth/Qwen3.6-27B-GGUF
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/cmogjhzqv0004jv04sl792yp8

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. --model
  3. Qwen3.6-27B-Q4_K_M.gguf
  4. --ctx-size
  5. 65536
  6. --n-gpu-layers
  7. 999
  8. --split-mode
  9. layer
  10. --tensor-split
  11. 0.5,0.5
  12. --cache-type-k
  13. q8_0
  14. --cache-type-v
  15. turbo4
  16. --batch-size
  17. 2048
  18. --ubatch-size
  19. 512
  20. --threads
  21. 8
FlagValue
--modelQwen3.6-27B-Q4_K_M.gguf
--ctx-size65536
--n-gpu-layers999
--split-modelayer
--tensor-split0.5,0.5
--cache-type-kq8_0
--cache-type-vturbo4
--batch-size2048
--ubatch-size512
--threads8

Measured speed

ConcurrencyContextPrefillDecodeTTFT msStatusSweep
165,536—37.8—observedqwen3-6-27b-q4-k-m-rtx-3090-24gb-llama-cpp-tp2-sl792yp8-sweep

Remaining fields

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

hardware count
2
hardware id
rtx-3090-24gb
id
qwen3-6-27b-q4-k-m-rtx-3090-24gb-llama-cpp-tp2-sl792yp8
model instance id
unsloth-qwen3-6-27b-gguf--q4-k-m
recipe source
localmaxxing
schema version
local-ai-registry/v1
speed sweep ids
qwen3-6-27b-q4-k-m-rtx-3090-24gb-llama-cpp-tp2-sl792yp8-sweep
status
candidate

capabilities

engine

name
llama.cpp
version
Tom-TurboQuant-fork-b3401eaf9

serving

max concurrency
1
max context tokens
65,536
tensor parallel
2
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
RTX 3090
localmaxxing · notes
Tom Turney's TurboQuant+ fork of llama.cpp. Asymmetric KV: q8_0 keys + turbo4 values per @no_stp_on_snek's twitter post. 2× RTX 3090 layer-split, 50/50 tensor-split. Pascal (2× 1080 Ti) got 14 TPS at 65K; Ampere gets 37.78 = 2.7× faster. CV 0.1% across 5 runs (essentially deterministic). Combined with the 35B-A3B Q3_K_XL bench (125.94 TPS), MoE-vs-dense advantage = 3.3× on Ampere (vs Pascal's 2.6×). Bench: 800-word essay × 1000 max_tokens × temp 0.6 streaming.
localmaxxing · observed command
llama-server --model Qwen3.6-27B-Q4_K_M.gguf --ctx-size 65536 --n-gpu-layers 999 --split-mode layer --tensor-split 0.5,0.5 --cache-type-k q8_0 --cache-type-v turbo4 --batch-size 2048 --ubatch-size 512 --threads 8
localmaxxing · run id
cmogjhzqv0004jv04sl792yp8
localmaxxing · tokenized · arguments
llama-server, --model, Qwen3.6-27B-Q4_K_M.gguf, --ctx-size, 65536, --n-gpu-layers, 999, --split-mode, layer, --tensor-split, 0.5,0.5, --cache-type-k, q8_0, --cache-type-v, turbo4, --batch-size, 2048, --ubatch-size, 512, --threads, 8
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/cmogjhzqv0004jv04sl792yp8 ↗