Local AI Registry

Recipe

gpt-oss-20b-mxfp4-rtx-3080-10gb-llama-cpp-tp1

gpt-oss-20b-mxfp4-rtx-3080-10gb-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
b10472
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/ggml-org/gpt-oss-20b-GGUF
Repository
ggml-org/gpt-oss-20b-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/cmt6lu7cs008mlk0129wtmtlb

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. gpt-oss-20b-MXFP4.gguf
  4. --host
  5. 127.0.0.1
  6. --port
  7. 8099
  8. --device
  9. CUDA0
  10. --n-gpu-layers
  11. 99
  12. --ctx-size
  13. 32768
  14. --flash-attn
  15. on
  16. --cache-type-k
  17. q8_0
  18. --cache-type-v
  19. q8_0
  20. --parallel
  21. 1
  22. --jinja
  23. --spec-type
  24. none
FlagValue
-mgpt-oss-20b-MXFP4.gguf
--host127.0.0.1
--port8099
--deviceCUDA0
--n-gpu-layers99
--ctx-size32768
--flash-attnon
--cache-type-kq8_0
--cache-type-vq8_0
--parallel1
--spec-typenone

Measured speed

ConcurrencyContextPrefillDecodeTTFT msStatusSweep
132,7681,46357.11,558.2observedgpt-oss-20b-mxfp4-rtx-3080-10gb-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-3080-10gb
id
gpt-oss-20b-mxfp4-rtx-3080-10gb-llama-cpp-tp1
model instance id
ggml-org-gpt-oss-20b-gguf--mxfp4
recipe source
localmaxxing
schema version
local-ai-registry/v1
speed sweep ids
gpt-oss-20b-mxfp4-rtx-3080-10gb-llama-cpp-tp1-sweep
status
candidate

capabilities

draft launch

accelerator backend
nvidia
arguments
-hf, ggml-org/gpt-oss-20b-GGUF, --n-gpu-layers, 999, --host, 0.0.0.0, --port, 8080, -c, 32768
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
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
RTX 3080
localmaxxing · notes
First RTX 3080 10GB entry for gpt-oss-20b. The 12.1 GB MXFP4 build does not fit a 10GB card outright, so it runs via --n-cpu-moe 9 (experts of the first 9 of 24 layers on CPU). TUNING NOTE: fewer offloaded layers is NOT automatically better. --n-cpu-moe 7 keeps more experts on the GPU but tips into host spill (574 MiB shared) and is worse on every axis: 43.9 tok/s decode, 404 tok/s prefill, 6190 ms TTFT. Backing off to 9 removes the spill (126 MiB shared) and gives 57.1 / 1463 / 1558 ms -- decode +30%, prefill 3.6x, TTFT 4x. VARIANCE: CPU-offloaded MoE decode is unstable here; five reps gave 62.2 / 32.9 / 61.7 / 50.3 / 57.1 tok/s, so treat 57.1 as a median rather than a repeatable figure. Also note ~1.2 GB of this card's VRAM is held by the Windows desktop (dwm + explorer) at idle, which moves the spill threshold between runs. Measured on an RTX 3080 10GB. Decode is the MEDIAN of 5 reps, not the best. Prompt is deliberately non-repetitive (repeated filler inflates decode on speculative setups by ~25%) and ignore_eos is set so output token count is identical across runs. Peak VRAM 9485 MiB with 126 MiB host-shared; elevated host-shared only indicates spill when the decode rate drops with it.
localmaxxing · observed command
llama-server \ -m gpt-oss-20b-MXFP4.gguf \ --host 127.0.0.1 --port 8099 \ --device CUDA0 --n-gpu-layers 99 \ --ctx-size 32768 \ --flash-attn on \ --cache-type-k q8_0 --cache-type-v q8_0 \ --parallel 1 --jinja \ --spec-type none
localmaxxing · run id
cmt6lu7cs008mlk0129wtmtlb
localmaxxing · tokenized · arguments
llama-server, -m, gpt-oss-20b-MXFP4.gguf, --host, 127.0.0.1, --port, 8099, --device, CUDA0, --n-gpu-layers, 99, --ctx-size, 32768, --flash-attn, on, --cache-type-k, q8_0, --cache-type-v, q8_0, --parallel, 1, --jinja, --spec-type, none
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/cmt6lu7cs008mlk0129wtmtlb ↗