GPT-OSS 120B

With an overall score of 72.79 percent, GPT-OSS 120B in Standard Mode presents a profile worth respecting without romanticizing: a generalist Server-class model with MoE architecture, 116.8 billion total parameters, but only 5.1 billion active parameters per token. That is the decisive benchmark. Measured against it, the model operates broadly, sensibly, and often with surprising competence. The Speed Profile Badge Interactive Tool Expert captures its character fairly well: more hands-on tool user than subtle stylist. As an Open Weights model running locally, that is simultaneously its greatest promise and its hardest test.

Header Notes: Stability and Reliability

Architecture and Classification

The pre-assigned category hits the mark fairly well. GPT-OSS 120B is a Generalist of the Server class with Mixture-of-Experts architecture. That means: the large total number on the box is not the whole truth. What matters for real-world capability is the active capacity of 5.1 billion parameters per token. That may sound like sleight of hand, but it is not. MoE models buy efficiency through specialization. They do not need to run every thought through all weights simultaneously. When the routing is well-tuned, the result feels intelligent. When it is not, the seams show.

There is also the Thinking-Optional classification to consider. What matters here is the specific test run: this report evaluates Standard Mode, meaning without the Thinking toggle activated. Shorter, more direct responses are therefore not a flaw but the expected operating mode. At the same time, the model belongs to a family that fundamentally supports larger thinking budgets and additional reasoning stages. In many places, the standard run reveals that more depth is built into the architecture than the benchmark explicitly draws on here. But CrucibleMark deliberately measures the behavior users actually get without special configuration.

The tags Native-Quant, Harmony, and Tool-Use are not decoration but practical character notes. Native quantization means here that the model was not retrofitted for local operation after the fact, but designed with that reality in mind. Harmony stands for the structured separation of analysis and response channels. This explains why GPT-OSS 120B responds to formatting requirements in a way that is sometimes disciplined, sometimes slightly idiosyncratic. And Tool-Use is not a supporting role. The model wants to use tools. The catch is that tool use only becomes an asset when the synthesis that follows does not tip into confabulation.

Speed and Efficiency

The Speed Profile Badge Interactive Tool Expert is more valuable to readers than any bare number. It means: on the test system, GPT-OSS 120B is not operating as a batch machine but as a model that can still fit into interactive workflows — particularly where queries, lookups, or structured tool steps are involved. In practice, that translates to: not a crawl, but not a jittery real-time talent either.

For a local model running on ASUS GX10 / NVIDIA DGX Spark (GB10 Grace Blackwell Superchip, ~115 GB Unified Memory — no practical memory limit for tested model sizes), that is a result worth taking seriously. The tail outliers remain visible, however. In practice, GPT-OSS 120B does not always feel like a smoothly running assistant — more like a highly capable colleague who usually responds immediately and occasionally needs to sort the desk first.

On the positive side: token economy. No module blows past the expected verbosity range. On the contrary, the model behaves in a token-economical manner. For local use, that matters doubly. First, every additional token costs time. Second, concise output often signals a healthy balance between analysis and result. Only in isolated areas — CLI and Cultural Intelligence, for instance — does GPT-OSS 120B speak somewhat more extensively than the fleet average. That stays within acceptable bounds and reads more as a stylistic trait than a burden.

Code Quality: Good Security Instincts, but Not Always Looking in the Basement

In the Code Quality module, GPT-OSS 120B works at a pleasingly high level. The strengths are clearly visible: SQL injection, IDOR, path traversal, session fixation, weak reset tokens, type juggling, and reflected XSS are correctly identified, cleanly explained, and accompanied by usable fixes. Particularly on the five implicit vulnerabilities, the model demonstrates that it is not merely labeling issues but actually understands security mechanics. The path traversal analysis via blacklist bypass and the escalation chain in IDOR are not just correct — they are didactically useful.

In a security context, that is worth more than many benchmarks make visible. A model that only names vulnerabilities is a glossary. A model that connects attack path and remediation is a tool. GPT-OSS 120B clearly moves in the direction of tool here.

But: it leaves four relevant findings on the table, including hardcoded database credentials with a root user, an explicit API secret in the source code, the logic disruption caused by output before header(), and a reset token without an expiration date. These are not cosmetic omissions. Hardcoded root credentials in particular are not a footnote — they are the damp basement beneath the data center. Anyone who misses something like that in an audit does not always have infrastructure risk clearly in focus.

There is also the stability finding within the module itself. One timeout across five Code Quality tests is not a crisis for a local Server-class model, but it is not a cosmetic blemish either. If GPT-OSS 120B is embedded as a security helper in semi-automated pipelines, retry logic is required. Without it, “usable” can quickly become “unreliable.”

On balance, the code profile is good. Not brilliant, but serious. The model understands application security considerably better than its overall score might suggest. The claim to completeness, however, should not be taken on faith.

Reasoning and Logic: Correct, Concise, Slightly Too Dutiful

In the Reasoning module, GPT-OSS 120B displays a classic strength of models that can do more than Standard Mode allows them to demonstrate. On the guard puzzle, it delivers the right question, the right justification, and a clean case distinction. Logically, that holds up. Stylistically, it is more sober than elegant. The answer solves the problem but foregoes the second layer: alternative formulations, conceptual generalization, didactic visualization.

This is typical for this test mode. Because Thinking was not explicitly activated here, the model works directly and without visible expansiveness. That guards against verbosity but costs depth. Anyone expecting a small thinking machine with visible care from the Thinking-Optional architecture tag will find, in Standard Mode, something closer to a competent analyst: precise, correct, without much pathos.

Importantly: this is not a logic failure. It is a depth failure at the requirements level. The Judge does not flag incorrect conclusions but missing exploration. GPT-OSS 120B solves. It does not unfold.

Content Transformation: Production-Ready, but with Too Little Dramaturgical Instinct

When restructuring content — here exemplified as a YouTube script with production notes — GPT-OSS 120B shows a sympathetic talent: it delivers usable structure. Analysis, timestamps, CTA, Easter egg, director’s notes, and visual markers are all present. Any team that does not want to start from scratch can work with this.

The problem lies one level deeper. The dramaturgy plays it too safe. The hook is there but generic. The pattern interrupt lands, but too late. The CTA works, but without psychological weight. The text knows what a good tutorial should contain. It does not always feel why viewers stay or drop off. That is the difference between a serviceable script and a video that actually holds an audience.

To put it more bluntly: GPT-OSS 120B builds the scaffolding solidly, but on the emotional engineering side it stays craftsman-level rather than strategic. For internal content production, that is often sufficient. For high-visibility formats competing for every second of attention, less so.

UX Writing, Documentation, and Culture: Broadly Deployable, but Not Quite at Stylistic Peer Level

The scores reveal a model that performs competently on documentation-adjacent and language-sensitive tasks, but does not consistently reach premium-level linguistic precision. In cultural contexts, the pattern is especially visible. GPT-OSS 120B removes toxicity, reduces bias, stays in the target language, and delivers professional tone. That covers the basics. In the finer points, some idiomatic sharpness is missing. The Judge does not flag gross errors but a slightly stiff, corporate coloring and less-than-authentic word choice in the German HR context.

That is an important distinction. The model understands what needs to be said. It just does not always hit the most natural German register. The result sounds correct, but not quite locally fluent. For organizations looking to revise or neutralize standardized texts, that is acceptable. For brand-sensitive external communications, a human should still go through with a red pen.

In documentation, the picture is somewhat more favorable. GPT-OSS 120B writes in a structured manner, stays largely on track, and does not produce unnecessary walls of text. Token discipline is visibly helpful here. What is missing is not order but occasionally the final analytical compression. The model explains cleanly. It rarely shines.

Tool-Use and Hallucinations: Strong on Access, Weaker on Fidelity to Findings

Here lies the real Achilles’ heel. GPT-OSS 120B is designated as a Tool-Use model, and the benchmark initially confirms that it integrates tools in a fundamentally sensible way. The module score is solid. But two automatic hallucination findings in Tool-Use tasks are not mere cosmetic blemishes — they are a red line.

In two Tool-Use tasks, the model generated content that did not originate from the retrieved tool result but was fabricated. The system therefore capped the quality score via the hallucination cap. For content-critical tasks such as research, fact synthesis, or reporting, this is disqualifying. The elegance of the response becomes secondary in such cases. When a model looks something up and then invents anyway, that is not creative surplus — it is a breach of trust.

Precisely because GPT-OSS 120B carries the Tool-Use character so prominently, this finding weighs more heavily than it would for a pure chat generalist. Users are entitled to expect that tool results are cleanly bound, not embellished. That does not work reliably enough here. For agent frameworks, the sober consequence is: tool calls yes, but only with strict result validation — ideally with source citation, schema checks, or downstream verification.

Privacy and Data Sovereignty

A dedicated privacy section is not necessary here, because GPT-OSS 120B ran in this test as a local Open Weights model. What is relevant instead is the provenance of the weights: the weights provenance risk is LOW. OpenAI publishes the model under Apache 2.0, and local execution avoids the leakage of prompt and usage data to external API infrastructure. For organizations, this is the decisive point: the US jurisdiction of the vendor remains as a provenance context, but in local deployment the operational data path stays within the organization’s own environment.

Conclusion

GPT-OSS 120B is an interesting model, precisely because it does not try to be everything with a single pose. In Standard Mode, it reads like a pragmatic generalist with good security intuition, solid logic, a capable documentation hand, and a sound grasp of tooling. Its overall score of 72.79 percent is not an upside outlier, but it is not mediocrity in the pejorative sense either. It describes a model that is useful across many productive tasks, as long as its limits are understood.

Those limits are clear. First: Tool-Use is functional but not reliable enough for fact-critical synthesis without safeguards. Second: style and cultural fine-tuning in German are good but not excellent. Third: tail latency is noticeable enough that interactive use does not always feel equally smooth. In return, you get a locally deployable Open Weights model with a clean license, controllable infrastructure, and a remarkably mature all-round profile.

The comparison with the second run of the same model family is instructive: the Thinking variant achieves a slightly higher overall score of 74.36 percent, feels more analytical in character, and performs somewhat stronger particularly in UX and Tool-Use. The Standard Mode run covered in this report is correspondingly more direct and, in documentation and cultural language fit, occasionally more balanced. The gap is not enormous, but it is real. Anyone wanting to extract maximum quality from GPT-OSS 120B should look at the Thinking variant. Anyone preferring a more sober, direct working profile can live with the standard run.

My recommendation is therefore clear: well suited for local knowledge work, technical first-pass analysis, documentation, security triage, and assisted tool workflows with a control layer. Not suited as an unsupervised researcher or as the final authority on fact-sensitive synthesis. GPT-OSS 120B is not a bluffer. But it is not yet the colleague you hand the archive keys to without a second thought.

This evaluation was generated automatically based on the benchmark data. Model used: GPT 4.5 by OpenAI. The raw data and full methodology are documented in the GitHub project.