MCMLV1
Published engineering benchmark · November 2026

GPT-OSS-120B on one heterogeneous workstation

Independently developed inference runtime built by Isaac Levinsky at MCMLV1 LLC. Benchmarked on the same workstation against tuned llama.cpp and ik_llama.cpp configurations at six prompt lengths.

NVIDIA Quadro GP100 16 GB + Quadro RTX 4000 8 GB · Intel i9-10900X · 128 GB system RAM
18 / 18Measured prefill, decode, and full-response comparisons won
2,113 / 2,113Reference next-token choices matched on the held-out suite
381Tokens/s prefill at 8K prompt length
17.3Tokens/s decode at 16K context

Read the benchmark report

Browse all ten pages below or download the original public PDF. The graphs cover prompt processing, answer generation, end-to-end latency, ongoing conversations, fidelity, and measurement method. The results are specific to the tested workstation and configurations.

Interactive reportPage 1 of 10
GPT-OSS-120B engineering benchmark cover slide, page 1 of 10
Click the page to open its full-size image. Use the arrows or thumbnails to navigate.

What was measured

Hardware-aware execution

A roughly 65 GB 117B-parameter MoE checkpoint across two generations of NVIDIA GPUs (24 GB combined VRAM), CPU compute, and system RAM.

Fair comparison

Source-built competitors with tuned settings, identical prompt tokens, 256-token greedy-decoded answers, and one request at a time.

Numerical fidelity

The custom runtime matched the original model's selected next token across 2,113 tested answer tokens. This is model fidelity, not a claim that model answers are always factually correct.

Scope of result

Measured on this one workstation from 336 to 65,280 prompt tokens. The public PDF discloses run counts; broader hardware claims would require separate benchmarking.

Engineering discussions

MCMLV1 works on inference runtimes, GPU kernels, numerical verification, memory movement, and private AI deployment. For a technical briefing or a conversation about paid engineering work, contact MCMLV1. Detailed benchmark records and technical method can be discussed under an appropriate NDA.