Evaluation record · gpt-oss-120b

GPT-OSS-120B

v20250805

OpenAI

Modelopen-sourceapache-2.0self-hostedprivacy

Exceptional

About This Model

OpenAI's first open-weight model released August 2025. 117B total params (5.1B active), Apache 2.0 license. Matches o4-mini on many benchmarks. Runs in 80GB memory.

Last Evaluated: July 9, 2026

Official Website

Trust Vector Analysis

Dimension Breakdown

🚀Performance & Reliability

Flagship open-source performance. MoE architecture activates 5.1B of 117B params per token. Matches or beats o4-mini on most benchmarks.

task accuracy code

Competition coding and tool use benchmarks

Evidence

Codeforces Benchmark — Matches o4-mini on competition coding

TauBench Tool Calling — Exceeds o4-mini on tool calling

highVerified: 2026-07-09

task accuracy reasoning

Math competition benchmarks

Evidence

AIME 2024 & 2025 — Outperforms o3-mini on competition mathematics

Chain-of-Thought Access — Full chain-of-thought reasoning process exposed

highVerified: 2026-07-09

task accuracy general

General knowledge and domain-specific testing

Evidence

MMLU & HLE — Matches o4-mini on general problem solving

HealthBench — Exceeds o4-mini on health-related queries

highVerified: 2026-07-09

output consistency

Internal testing

Evidence

OpenAI Model Card — Configurable reasoning effort for consistency

mediumVerified: 2026-07-09

latency p50

Median latency estimation

Evidence

Optimized Inference — Fast inference with MoE architecture

mediumVerified: 2026-07-09

latency p95

95th percentile from community benchmarks

Evidence

Community Deployments — ~2s on H100 hardware

mediumVerified: 2026-07-09

context window

Official specification

Evidence

OpenAI Technical Specs — 128K context window natively supported

OpenAI Model Docs — 131,072-token context window and up to 131,072 max output tokens

highVerified: 2026-07-09

uptime

Self-hosting provides full control

Evidence

Self-Hosted Model — 100% uptime when self-hosted

highVerified: 2026-07-09

🛡️Security

Good base security. Self-hosting provides complete control over safety guardrails and data handling. Note: Feb 2026 research (arXiv 2602.14689) shows open-weight models broadly remain vulnerable to prefill attacks — pair self-hosted deployments with external guardrails. OpenAI's gpt-oss-safeguard (Oct 2025), an Apache-2.0 safety-classifier fine-tune of gpt-oss, can serve as a policy-based moderation layer.

prompt injection resistance

OWASP LLM01 testing

Evidence

OpenAI Safety Testing — Good resistance, customizable for self-hosted

mediumVerified: 2026-07-09

jailbreak resistance

Adversarial testing

Evidence

Community Testing — Standard resistance, self-host allows custom guardrails

Prefill Attack Study (arXiv 2602.14689) — Large empirical study finds prefill attacks consistently effective against all major contemporary open-weight models; large reasoning models show partial resistance but remain vulnerable to tailored strategies

mediumVerified: 2026-07-09

data leakage prevention

Self-hosting analysis

Evidence

Self-Hosted Deployment — Complete data control when self-hosted

highVerified: 2026-07-09

output safety

Safety testing

Evidence

OpenAI Safety — Standard safety training, customizable

mediumVerified: 2026-07-09

api security

Deployment security review

Evidence

Self-Hosted Security — Customer controls all API security when self-hosted

highVerified: 2026-07-09

🔒Privacy & Compliance

Perfect privacy when self-hosted. No data sent to OpenAI. Full compliance control. Ideal for regulated industries.

data residency

Self-hosting analysis

Evidence

Open-Weight Model — Deploy anywhere, full data residency control

highVerified: 2026-07-09

training data optout

Privacy model analysis

Evidence

Self-Hosted Model — No data sent to OpenAI when self-hosted

highVerified: 2026-07-09

data retention

Self-hosting review

Evidence

Self-Hosted Deployment — Complete control over data retention

highVerified: 2026-07-09

pii handling

Data flow analysis

Evidence

On-Premises Deployment — PII never leaves your infrastructure

highVerified: 2026-07-09

compliance certifications

Compliance model review

Evidence

Self-Hosted Compliance — Inherit your infrastructure's certifications

highVerified: 2026-07-09

zero data retention

Privacy architecture review

Evidence

Open-Weight Model — Complete control, zero external retention

highVerified: 2026-07-09

👁️Trust & Transparency

Exceptional transparency. Full chain-of-thought access. Complete model weights and architecture disclosed. Open-source enables auditing.

explainability

Reasoning transparency

Evidence

Full Chain-of-Thought — Complete access to reasoning process

highVerified: 2026-07-09

hallucination rate

QA testing

Evidence

Benchmark Testing — Good factual accuracy

mediumVerified: 2026-07-09

bias fairness

Bias benchmarks

Evidence

OpenAI Model Card — Standard bias testing

mediumVerified: 2026-07-09

uncertainty quantification

Confidence assessment

Evidence

Model Behavior — Good uncertainty expression

mediumVerified: 2026-07-09

model card quality

Documentation review

Evidence

Comprehensive Model Card — Detailed technical specs, benchmarks, architecture

highVerified: 2026-07-09

training data transparency

Training data disclosure review

Evidence

OpenAI Documentation — Mostly English, STEM, coding focus disclosed

highVerified: 2026-07-09

guardrails

Safety mechanism review

Evidence

Customizable Guardrails — Standard safety, customizable when self-hosted

highVerified: 2026-07-09

⚙️Operational Excellence

Exceptional operational flexibility. Apache 2.0 enables commercial use. Massive deployment ecosystem. Self-host or use managed platforms.

api design quality

API compatibility review

Evidence

Deployment Platforms — Works with vLLM, Ollama, llama.cpp, Azure, AWS, etc.

highVerified: 2026-07-09

sdk quality

SDK ecosystem review

Evidence

GitHub Repository — Official repo, Hugging Face integration

highVerified: 2026-07-09

versioning policy

Version stability analysis

Evidence

Open Weights — Weights frozen, no deprecation risk

Hugging Face Repository — Repo last updated 2025-08-26; weights stable with no new revisions, remains freely downloadable under Apache 2.0

highVerified: 2026-07-09

monitoring observability

Monitoring capability review

Evidence

Self-Hosted Control — Full observability when self-hosted

highVerified: 2026-07-09

support quality

Support ecosystem assessment

Evidence

Community Support — GitHub issues, community forums, deployment partners

mediumVerified: 2026-07-09

ecosystem maturity

Ecosystem breadth analysis

Evidence

Deployment Partners — Azure, Hugging Face, AWS, Fireworks, Together AI, Databricks, Vercel, Cloudflare, OpenRouter

highVerified: 2026-07-09

license terms

License review

Evidence

Apache 2.0 License — Permissive Apache 2.0, no copyleft, no patent risk

highVerified: 2026-07-09

Strengths

+Apache 2.0 open-weight license enables commercial use without restrictions
+Matches or beats o4-mini on coding, math, and health benchmarks
+Complete data privacy when self-hosted (zero external data transmission)
+Full chain-of-thought reasoning access for transparency and debugging
+MoE architecture: 5.1B active of 117B total params, runs in 80GB
+Massive deployment ecosystem (Azure, AWS, Hugging Face, vLLM, Ollama)

Limitations

!Requires 80GB GPU memory (H100 or equivalent)
!Self-hosting complexity and infrastructure costs
!Community support vs enterprise SLA
!Slightly lower performance than flagship closed models
!No built-in safety guardrails (customizable but requires setup)

Metadata

pricing

input: Free (self-hosted)

output: Free (self-hosted)

notes: Infrastructure costs only: ~$2-4/hr for H100. Managed platforms vary. Free for download and commercial use under Apache 2.0.

last verified: 2026-07-09

context window: 131072

languages

0: English

1: Spanish

2: French

3: German

4: Italian

5: Portuguese

6: Japanese

7: Korean

8: Chinese

modalities

0: text

api endpoint: Self-hosted (various platforms)

model download: https://huggingface.co/openai/gpt-oss-120b

github: https://github.com/openai/gpt-oss

open source: true

license: Apache 2.0

architecture: Mixture-of-Experts (MoE) Transformer

parameters: 117B total (5.1B active per token)

memory requirement: 80GB (MXFP4 quantization)

tokenizer: o200k_harmony

deployment platforms

0: Azure

1: AWS

2: Hugging Face

3: vLLM

4: Ollama

5: llama.cpp

6: LM Studio

7: Fireworks

8: Together AI

9: Baseten

10: Databricks

11: Vercel

12: Cloudflare

13: OpenRouter

Use Case Ratings

code generation

Excellent coding. Matches o4-mini. Configurable reasoning effort. Full chain-of-thought debugging.

customer support

Good for customer support. Self-host for complete data privacy. Configurable reasoning for cost control.

content creation

Strong content creation. Self-hosting enables unlimited generation without API costs.

data analysis

Excellent for data analysis. Keep sensitive data on-premises. Full chain-of-thought for transparency.

research assistant

Outstanding for research. 128K context. Self-host proprietary research data. Full reasoning transparency.

legal compliance

Perfect for legal. Self-host for complete compliance. No data leaves premises. Apache 2.0 license clarity.

healthcare

Ideal for healthcare. Self-host for HIPAA. Complete PHI privacy. No external data transmission.

financial analysis

Excellent for finance. Outperforms o3-mini on math. Self-host proprietary financial data.

education

Great for education. Full chain-of-thought shows reasoning steps. Self-host for institutional control.

creative writing

Good creative writing. Unlimited generation when self-hosted. No API costs for iteration.

Similar Models

GPT-OSS-20B

OpenAI

OpenAI o4-mini

OpenAI

OpenAI o3-mini

OpenAI

Claude Haiku 4.5

Anthropic