Local AI Registry

Recipe

qwen3-6-27b-iq4-xs-rtx-3090-24gb-llama-cpp-tp1

qwen3-6-27b-iq4-xs-rtx-3090-24gb-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
am17an/mtp (upstream PR #22673)
Accelerators
1
Tensor parallel
1
Context tokens
131,072
Max concurrency
1
chat
unknown
reasoning
unknown
tools
unknown
vision
unknown

Hugging Face model card

Identity

https://huggingface.co/unsloth/Qwen3.6-27B-MTP-GGUF
Repository
unsloth/Qwen3.6-27B-MTP-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/cmpa2l8j704qho401hmhgyeoe

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-cpp-mtp/build/bin/llama-server
  2. --prio
  3. 3
  4. --ctx-size
  5. 131072
  6. --flash-attn
  7. on
  8. --cache-type-k
  9. q4_0
  10. --cache-type-v
  11. q4_0
  12. --no-mmap
  13. --n-gpu-layers
  14. auto
  15. --model
  16. /mnt/g/models/unsloth/Qwen3.6-27B-MTP-GGUF/Qwen3.6-27B-IQ4_XS.gguf
  17. --offline
  18. --temperature
  19. 1.0
  20. --top-k
  21. 20
  22. --top-p
  23. 0.95
  24. --min-p
  25. 0.0
  26. --repeat-penalty
  27. 1.0
  28. --presence-penalty
  29. 1.5
  30. --ctx-checkpoints
  31. 4
  32. --checkpoint-every-n-tokens
  33. 8192
  34. --cache-ram
  35. 4096
  36. --kv-unified
  37. --parallel
  38. 1
  39. --mmproj
  40. /mnt/g/models/unsloth/Qwen3.6-27B-MTP-GGUF/mmproj-BF16.gguf
  41. --no-mmproj-offload
  42. --image-min-tokens
  43. 1024
  44. --reasoning
  45. on
  46. --chat-template-kwargs
  47. {"preserve_thinking":true}
  48. --reasoning-budget
  49. 8192
  50. --spec-type
  51. draft-mtp
  52. --spec-draft-n-max
  53. 2
  54. --spec-draft-p-min
  55. 0.50
  56. --spec-draft-n-min
  57. 2
FlagValue
--prio3
--ctx-size131072
--flash-attnon
--cache-type-kq4_0
--cache-type-vq4_0
--n-gpu-layersauto
--model/mnt/g/models/unsloth/Qwen3.6-27B-MTP-GGUF/Qwen3.6-27B-IQ4_XS.gguf
--temperature1.0
--top-k20
--top-p0.95
--min-p0.0
--repeat-penalty1.0
--presence-penalty1.5
--ctx-checkpoints4
--checkpoint-every-n-tokens8192
--cache-ram4096
--parallel1
--mmproj/mnt/g/models/unsloth/Qwen3.6-27B-MTP-GGUF/mmproj-BF16.gguf
--image-min-tokens1024
--reasoningon
--chat-template-kwargs{"preserve_thinking":true}
--reasoning-budget8192
--spec-typedraft-mtp
--spec-draft-n-max2
--spec-draft-p-min0.50
--spec-draft-n-min2

Measured speed

ConcurrencyContextPrefillDecodeTTFT msStatusSweep
1131,072—40.8—observedqwen3-6-27b-iq4-xs-rtx-3090-24gb-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
rtx-3090-24gb
id
qwen3-6-27b-iq4-xs-rtx-3090-24gb-llama-cpp-tp1
model instance id
unsloth-qwen3-6-27b-mtp-gguf--iq4-xs
recipe source
localmaxxing
schema version
local-ai-registry/v1
speed sweep ids
qwen3-6-27b-iq4-xs-rtx-3090-24gb-llama-cpp-tp1-sweep
status
candidate

capabilities

draft launch

accelerator backend
nvidia
arguments
-hf, unsloth/Qwen3.6-27B-MTP-GGUF:IQ4_XS, --n-gpu-layers, 999, --host, 0.0.0.0, --port, 8080, -c, 131072
container port
8,080
environment · LLAMA CACHE
/root/.cache/huggingface
host port
8,080
image
ghcr.io/ggml-org/llama.cpp:server-cuda12-b10481@sha256:b2497f8834f5ecb4e38530f6bf2734b8e0be107ff48e4720145911c86930f2ce
ipc
host
kind
docker
shm size
16g
synthesized · generated at
2026-08-31T22:12:17Z
synthesized · image provenance
gemma-4-12b-q4-k-m-rtx-3060-12gb-llama-cpp-tp1
synthesized · template
llama-cpp-server-v1

mounts

read onlytarget
No/root/.cache/huggingface

engine

name
llama.cpp
version
am17an/mtp (upstream PR #22673)

serving

max concurrency
1
max context tokens
131,072
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
draft launch.entrypoint · provenance · captured at
2026-09-01T01:41:26Z
draft launch.entrypoint · reason linux-amd64-container-config-entrypoint
draft launch.entrypoint · state known
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
serving.max concurrency · provenance · captured at
2026-08-31T23:03:15Z
serving.max concurrency · reason server-capacity-derived-from-source-evidence
serving.max concurrency · state known

metadata

localmaxxing · backend
cuda
localmaxxing · batch size
2,048
localmaxxing · hardware label
RTX 3090
localmaxxing · notes
MTP (Multi-Token Prediction) benchmark on unsloth/Qwen3.6-27B-MTP-GGUF IQ4_XS. Full grid search of 16 configs (n_max: 2/4/6/8/16 × p_min: 0.35/0.50/0.75 + baseline). Best: --spec-draft-n-max 2 --spec-draft-p-min 0.50 at 40.8 tok/s (1.47x vs 27.8 baseline no-MTP). n_max=16 was slower than baseline (27.0 tok/s). llama.cpp upstream PR #22673, am17an/mtp branch. Server restarted between each config. 256 output tokens, single request, temperature=1.0 (non-greedy). MTP acceptance rate ~75% at n_max=2. Build commit: upstream llama.cpp from mid-May 2026.
localmaxxing · observed command
~/llama-cpp-mtp/build/bin/llama-server --prio 3 --ctx-size 131072 --flash-attn on --cache-type-k q4_0 --cache-type-v q4_0 --no-mmap --n-gpu-layers auto --model /mnt/g/models/unsloth/Qwen3.6-27B-MTP-GGUF/Qwen3.6-27B-IQ4_XS.gguf --offline --temperature 1.0 --top-k 20 --top-p 0.95 --min-p 0.0 --repeat-penalty 1.0 --presence-penalty 1.5 --ctx-checkpoints 4 --checkpoint-every-n-tokens 8192 --cache-ram 4096 --kv-unified --parallel 1 --mmproj /mnt/g/models/unsloth/Qwen3.6-27B-MTP-GGUF/mmproj-BF16.gguf --no-mmproj-offload --image-min-tokens 1024 --reasoning on --chat-template-kwargs '{"preserve_thinking":true}' --reasoning-budget 8192 --spec-type draft-mtp --spec-draft-n-max 2 --spec-draft-p-min 0.50 --spec-draft-n-min 2
localmaxxing · run id
cmpa2l8j704qho401hmhgyeoe
localmaxxing · tokenized · arguments
~/llama-cpp-mtp/build/bin/llama-server, --prio, 3, --ctx-size, 131072, --flash-attn, on, --cache-type-k, q4_0, --cache-type-v, q4_0, --no-mmap, --n-gpu-layers, auto, --model, /mnt/g/models/unsloth/Qwen3.6-27B-MTP-GGUF/Qwen3.6-27B-IQ4_XS.gguf, --offline, --temperature, 1.0, --top-k, 20, --top-p, 0.95, --min-p, 0.0, --repeat-penalty, 1.0, --presence-penalty, 1.5, --ctx-checkpoints, 4, --checkpoint-every-n-tokens, 8192, --cache-ram, 4096, --kv-unified, --parallel, 1, --mmproj, /mnt/g/models/unsloth/Qwen3.6-27B-MTP-GGUF/mmproj-BF16.gguf, --no-mmproj-offload, --image-min-tokens, 1024, --reasoning, on, --chat-template-kwargs, {"preserve_thinking":true}, --reasoning-budget, 8192, --spec-type, draft-mtp, --spec-draft-n-max, 2, --spec-draft-p-min, 0.50, --spec-draft-n-min, 2
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/cmpa2l8j704qho401hmhgyeoe ↗