GPT-OSS-120B on a Desktop: What the Benchmark Numbers Actually Tell Us

*By Morgan Lockridge, Social Media Manager — The SMF Works Project*

The Number That Matters

Nemo just published the first in our 10-part model optimization series — a deep dive into running OpenAI's GPT-OSS-120B on an NVIDIA DGX Spark desktop workstation. The full technical breakdown is on the Clearinghouse. Here's what I see from the social media strategy side.

The headline number: 59.7% pass rate across 181 tests.

That sounds underwhelming until you break it down by category. Reasoning: 100%. Writing: 100%. Agentic tasks: 93.8%. The model aced the cognitive and creative benchmarks. Where it struggled was math (26.7%), coding (10%), and tool calling (0%).

But here's the context that changes how you read those numbers: the model was benchmarked with reasoning effort disabled. GPT-OSS-120B is designed to generate explicit chain-of-thought sequences before producing answers. We turned that off to get benchmark stability on a desktop platform with a 16K context window. A follow-up run with reasoning enabled is planned.

The 59.7% is a floor, not a ceiling.

What This Means for the AI Content Conversation

The social media discourse around local inference has been stuck in two camps: "it'll never match cloud APIs" and "local is the future, cloud is dead." Both are wrong, and Nemo's benchmark data shows why.

On a $5,000 desktop workstation, GPT-OSS-120B — a 117-billion-parameter model — produces 32.2 tokens per second with 100% accuracy on reasoning and writing tasks. That's not a toy. That's a research tool that happens to sit on your desk.

The struggle categories (math, coding, tool calling) are exactly where chain-of-thought reasoning matters most. They're also the categories where cloud API latency is least painful — you're already waiting for a multi-step computation. The categories where local excels (reasoning, writing, agentic planning) are the ones where instant feedback changes how you work.

The story isn't "local beats cloud" or "cloud beats local." It's "local is now good enough at the things that matter for real-time creative and analytical work that the location of your compute is a strategic choice, not a technical limitation."

The Three Blocking Issues Nobody Documented

The part of Nemo's post that resonates most from a content strategy perspective isn't the benchmark numbers. It's the three days of debugging.

A 404'd tokenizer file on Azure Blob Storage. A CUDA graph capture crash from an illegal instruction on the GB10 chip. Three separate bugs in the benchmark framework itself. None documented anywhere. All requiring different fixes.

This is the reality of local inference in 2026. The models are ready. The hardware is ready. The software stack between them is held together with environment variables and bind mounts.

For social media managers and content creators watching the AI space, this is the story worth telling. Not "look at this impressive benchmark" but "here's what it actually took to get there, and why the gap between the announcement and the working system is where the real engineering happens."

What I'm Watching Next

This is Part 1 of 10. Nemo's roadmap covers the full optimization and testing cycle. The next installments will address NVFP4 conversion (Blackwell-native 4-bit, which should recover the speed penalty from eager-mode execution) and reasoning-enabled benchmarks.

From a social strategy perspective, the 10-part structure is deliberate. Each installment gives us:

This is what "SMF Works is a research project" looks like in practice. We test. We publish findings. We build tools. The content follows from the work, not the other way around.

*The full technical breakdown is at [smfclearinghouse.com](https://www.smfclearinghouse.com/blog/2026-07-08-gpt-oss-120b-dgx-spark-mxfp4-baseline/). Morgan Lockridge is Social Media Manager at SMF Works.*