Model benchmarks

Ornith-1.0-35B-MTPLX local LLM performance

As of August 2026, Ornith-1.0-35B-MTPLX runs at up to 124.0 tok/s for local inference (best of 2 community benchmark runs across 1 GPU).

oMLX
ShareRedditX

Model size

35B

Peak speed

124.0 tok/s

Average speed

104.5 tok/s

Min memory

18.9 GB

Max context

65,536 tokens

Avg runtime / run

2m 7s

Avg quality

80.3

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)
Overall79.380.381.3
Agent Workflow79.281.984.7
Code Generation68.670.772.8
Role Play & Narrative90.090.190.3
Research & Analysis77.878.479.1

Performance by hardware and tool

Every hardware/tool/quantization combination Ornith-1.0-35B-MTPLX has been benchmarked on, ranked by peak token generation speed. Last updated August 2026.

HardwareToolQuantBest tok/sAvg tok/sMemoryContextQualityRuns
Apple M5 MaxoMLX—124.0 tok/s104.5 tok/s18.9 GB65,536 tokens80.32

Benchmark runs

All 2 Ornith-1.0-35B-MTPLX runs submitted so far — expand one for its hardware, quality breakdown and per-scenario detail.

Frequently asked questions

Is Ornith-1.0-35B-MTPLX good for agentic (tool-using) tasks?
In our benchmarks, Ornith-1.0-35B-MTPLX scores 81.9/100 for agentic workflows. It runs at about 104.5 tok/s, so if you want more speed, Tiel-Coder-35B-A3B-MLX-oQ4e-MTP is faster (~123.7 tok/s) and still scores well for agentic workflows (84.4/100).
How fast is Ornith-1.0-35B-MTPLX for local inference?
Across 2 community benchmark runs, Ornith-1.0-35B-MTPLX reaches up to 124.0 tok/s and averages 104.5 tok/s, with the fastest results on Apple M5 Max.
How much memory does Ornith-1.0-35B-MTPLX need?
The leanest observed configuration used about 18.9 GB of memory.
Which tools have been used to run Ornith-1.0-35B-MTPLX?
Benchmarks were submitted using oMLX. Results are community-contributed and updated as new runs arrive.