Model benchmarks

ox253 local LLM performance

As of November 2026, ox253 runs at up to 56.9 tok/s for local inference (best of 2 community benchmark runs across 1 GPU).

llama.cpp
ShareRedditX

Model size

Unknown

Peak speed

56.9 tok/s

Average speed

37.7 tok/s

Avg PP

732.6 tok/s

Min memory

n/a

Max context

65,536 tokens

Avg output / run

22,618 tokens

Avg runtime / run

10m 30s

Avg quality

64.6

Benchmark runs

2

GPUs tested

1

Quality by task

Average LLM-judged quality (0–100) with the run-to-run spread shown as a P5–P95 band, overall and for each benchmark task, across all 2 runs. The low and high columns show how much the judge’s score varies between runs, and need at least two runs to display.

TaskP5 (low)AvgP95 (high)
Overall56.964.672.2
Agent Workflow20.547.674.6
Code Generation65.266.668.0
Role Play & Narrative74.077.180.1
Research & Analysis65.267.169.0

Performance by hardware and tool

Every hardware/tool/quantization combination ox253 has been benchmarked on, ranked by peak token generation speed. Last updated November 2026.

HardwareToolQuantBest tok/sAvg tok/sMemoryContextQualityRuns
Quadro K620llama.cpp—56.9 tok/s37.7 tok/sn/a65,536 tokens64.62

Benchmark runs

All 2 ox253 runs submitted so far — expand one for its hardware, quality breakdown and per-scenario detail.

Frequently asked questions

Is ox253 good for coding?
In our benchmarks, ox253 scores 66.6/100 for coding. It runs at about 37.7 tok/s, so if you want more speed, turbo is faster (~124.1 tok/s) and still scores well for coding (85.3/100).
Is ox253 good for agentic (tool-using) tasks?
In our benchmarks, ox253 scores 47.6/100 for agentic workflows. It runs at about 37.7 tok/s, so if you want more speed, granite-4.2-8b-Q4_K_L is faster (~71.6 tok/s) and still scores well for agentic workflows (92.9/100).
How fast is ox253 for local inference?
Across 2 community benchmark runs, ox253 reaches up to 56.9 tok/s and averages 37.7 tok/s, with the fastest results on Quadro K620.
Which tools have been used to run ox253?
Benchmarks were submitted using llama.cpp. Results are community-contributed and updated as new runs arrive.