Recipe

qwen3-8-27b-iq2-xxs-rtx-3080-10gb-llama-cpp-tp1-hfssqgv8

qwen3-8-27b-iq2-xxs-rtx-3080-10gb-llama-cpp-tp1-hfssqgv8

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

Record

Status
candidate
Source
localmaxxing
Engine
llama.cpp
Engine version
b10472
Accelerators
1
Tensor parallel
1
Context tokens
16,384
Max concurrency
1
chat
unknown
reasoning
unknown
tools
unknown
vision
unknown

Hugging Face model card

Identity

https://huggingface.co/unsloth/Qwen3.8-27B-GGUF
Repository
unsloth/Qwen3.8-27B-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/cmsy6c9rz0crzms01hfssqgv8

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. -m
  3. Qwen3.8-27B-UD-IQ2_XXS.gguf
  4. -c
  5. 16384
  6. -ngl
  7. 99
  8. -fa
  9. on
  10. -ctk
  11. q8_0
  12. -ctv
  13. q8_0
  14. --parallel
  15. 1
  16. --jinja
  17. --spec-type
  18. draft-mtp
  19. --spec-draft-n-max
  20. 2
FlagValue
-mQwen3.8-27B-UD-IQ2_XXS.gguf
-c16384
-ngl99
-faon
-ctkq8_0
-ctvq8_0
--parallel1
--spec-typedraft-mtp
--spec-draft-n-max2

Measured speed

ConcurrencyContextPrefillDecodeTTFT msStatusSweep
116,3841,044.270.32,581.2observedqwen3-8-27b-iq2-xxs-rtx-3080-10gb-llama-cpp-tp1-hfssqgv8-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-3080-10gb
id
qwen3-8-27b-iq2-xxs-rtx-3080-10gb-llama-cpp-tp1-hfssqgv8
model instance id
unsloth-qwen3-8-27b-gguf--iq2-xxs
recipe source
localmaxxing
schema version
local-ai-registry/v1
speed sweep ids
qwen3-8-27b-iq2-xxs-rtx-3080-10gb-llama-cpp-tp1-hfssqgv8-sweep
status
candidate

capabilities

draft launch

accelerator backend
nvidia
arguments
-hf, unsloth/Qwen3.8-27B-GGUF:IQ2_XXS, --n-gpu-layers, 999, --host, 0.0.0.0, --port, 8080, -c, 16384
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:b2497f8834f5ecb4e38530f6bf2734b8e0be107f0f0857e259672d1cb85b71c2
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
b10472

serving

max concurrency
1
max context tokens
16,384
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
RTX 3080
localmaxxing · notes
MTP arm of a paired A/B (+58.3% over the spec-off row). Draft acceptance 0.796 (156/196), identical across all 3 reps. Headless card (both displays moved to the 7900X iGPU; desktop holds 64 MiB on the 3080 - with displays attached dwm holds 1846 MiB, full offload does not fit, and baseline drops to 33.0). Full offload, every layer on GPU. Workload: single 2700-token non-repetitive synthetic log-summarisation prompt, 256 output tokens, temp 0, cache_prompt off; medians of 3 reps. Prefill/decode from llama-server timings, TTFT measured client-side on a streaming request. NOTE ON WORKLOAD SENSITIVITY: a repeated-sentence prompt drives MTP acceptance to 0.98 and inflates decode to 81.4 tok/s; this prompt is deliberately non-repetitive. Cross-check with the qwen38-mtp repo's stock probe.py (3 diverse code/prose prompts, medians of 3 complete passes) gives 45.1 spec-off and 64.4 with MTP n-max 2 (+43%) at 0.74 acceptance - lower than this row's +58% because acceptance is workload-dependent. Context ceiling on this card: 16K is the largest window where the MTP arm stays fully resident; at 32K it spills to host RAM over PCIe and inverts to ~25 tok/s.
localmaxxing · observed command
llama-server -m Qwen3.8-27B-UD-IQ2_XXS.gguf -c 16384 -ngl 99 -fa on -ctk q8_0 -ctv q8_0 --parallel 1 --jinja --spec-type draft-mtp --spec-draft-n-max 2
localmaxxing · run id
cmsy6c9rz0crzms01hfssqgv8
localmaxxing · tokenized · arguments
llama-server, -m, Qwen3.8-27B-UD-IQ2_XXS.gguf, -c, 16384, -ngl, 99, -fa, on, -ctk, q8_0, -ctv, q8_0, --parallel, 1, --jinja, --spec-type, draft-mtp, --spec-draft-n-max, 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/cmsy6c9rz0crzms01hfssqgv8