LLMPrivate

GPT OSS 120B

OpenAI's largest open-weight reasoning model — 117B MoE parameters, Apache 2.0 licensed, with web search and full chain-of-thought.

Maker
OpenAI
Modality
Text
Context
128K tokens
Open weights
Yes — Apache 2.0

Overview

What is GPT OSS 120B

GPT OSS 120B is OpenAI's 117B-parameter open-weight reasoning model released in August 2025 under the Apache 2.0 license. It supports configurable reasoning effort, full chain-of-thought, native web search, and efficient deployment on a single 80GB GPU for data sovereignty.

Running it privately on Venice

On Venice, GPT OSS 120B runs inside a trusted execution environment with end-to-end encryption and zero retention — your prompts are never stored, profiled, or used for training. You get the same open-weight model with reasoning and web-search capabilities, but with full data sovereignty and no Big-Tech surveillance.

Private (zero retention)No prompt trainingTEE · hardware enclaveEnd-to-end encrypted

Assessment

Strengths and limitations

Strengths
  • Open weights under Apache 2.0: permissionless self-hosting, fine-tuning, and commercial use without copyleft or patent risk.
  • Configurable reasoning effort (low, medium, high) with complete chain-of-thought visibility for debugging and trust.
  • Native web search and reasoning on Venice, with broader agentic capabilities available when self-hosted.
  • Efficient deployment: fits on a single 80GB GPU (H100 or AMD MI300X) despite 117B parameters thanks to MoE architecture and MXFP4 quantization.
  • Strong instruction following designed for agentic workflows and structured outputs.
Limitations
  • Text-only: no vision, image, or audio support.
  • Requires the harmony response format; using other formats will cause incorrect behavior.
  • Open-weight release means determined attackers can fine-tune away safety refusals, so downstream systems need extra guardrails.
  • Not uncensored: follows OpenAI safety policies by default, which may refuse certain requests.
  • The smaller GPT OSS 20B can be more efficient for latency-sensitive tasks and sometimes matches or exceeds this model on select benchmarks.

Capabilities

What it supports

  • Tool use / function calling
  • Vision (image input)
  • Reasoning
  • Web search
  • Code-optimized
  • Structured output (JSON schema)
  • Audio input
  • Video input
  • Multiple image inputs
  • Log probabilities

Specifications

Datasheet

Maker
OpenAI
Released
August 5, 2025
Architecture
Mixture of Experts (MoE)
Parameters
117B total / 5.1B active
Open weights
Yes — Apache 2.0
Context window
128K tokens
Max output
32.768K tokens
Capabilities
Reasoning, Web search
Privacy on Venice
Private — zero retention
Available on Venice since
Mar 2026

API

Call it from your code

Venice exposes an OpenAI-compatible API. Point your base URL at Venice and pass the model id.

curl https://api.venice.ai/api/v1/chat/completions \
  -H "Authorization: Bearer $VENICE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "e2ee-gpt-oss-120b-p",
    "messages": [{ "role": "user", "content": "Explain quantum tunneling simply." }]
  }'

Pricing

What it costs on Venice

Billed per token on Venice: $0.13 per 1M input tokens and $0.65 per 1M output tokens.

Input / 1M tokens
$0.13
Per 1M tokens
Output / 1M tokens
$0.65
Per 1M tokens

New Venice accounts include a free daily allowance and 500 welcome credits — no credit card required.

Alternatives

How it compares

ModelContext windowOpen weightsPrice (Venice)Best for
GPT OSS 120B128K tokensYes$0.13 in · $0.65 out / 1MOpen-weight reasoning & web search
DeepSeek V3.2160K tokensYes$0.33 in · $0.48 out / 1MLong-context open MoE
Kimi K2.6256K tokensYes$0.75 in · $3.50 out / 1MLong-context reasoning
Claude Sonnet 4.61M tokensNo$3.60 in · $18 out / 1MClosed frontier agent

OpenAI's most powerful open model — full CoT, Apache 2.0, and efficient on a single GPU.

Use cases

What it is good for

  1. 01Self-hosted or on-premise reasoning for enterprises that need data sovereignty and zero retention.
  2. 02Agentic workflows with web search, structured outputs, and chain-of-thought debugging.
  3. 03Fine-tuning for domain-specific instruction following or compliance requirements.
  4. 04Cost-efficient large-scale inference where open-weight control and permissionless deployment matter.
  5. 05Privacy-sensitive research and development via Venice's encrypted, zero-retention inference.

Prompting

Getting better results

Always use the harmony response format — the model is trained exclusively for it and will malfunction with other formats.

Explicitly set reasoning effort (low, medium, high) to balance latency against answer depth.

Request structured outputs when you need machine-parseable results.

Inspect chain-of-thought for debugging, but avoid exposing raw reasoning traces to end users.

Version history

GPT OSS 20B
2025-08

Smaller sibling — 21B parameters, faster, edge-device friendly.

GPT OSS 120B
2025-08

Current — largest open-weight model in the series.

FAQ

Frequently asked questions

GPT OSS 120B is OpenAI's largest open-weight language model, released in August 2025 under the Apache 2.0 license. It has 117B total parameters with 5.1B active per token, supports reasoning, web search, and full chain-of-thought, and fits on a single 80GB GPU.

Venice bills per token at $0.13 per 1M input tokens and $0.65 per 1M output tokens. There is no subscription required — you only pay for what you use.

It is open weights under the Apache 2.0 license, so you can download, self-host, and fine-tune the model. However, the training data and full training code are not open source.

Yes. On Venice it exposes native web search and reasoning. Because the weights are open, you can also self-host to enable function calling, Python execution, and structured outputs via the OpenAI-compatible API.

On Venice, GPT OSS 120B runs in a trusted execution environment with end-to-end encryption and zero retention. Your prompts are not stored, profiled, or used for training.

Choose GPT OSS 120B for cheaper input pricing, full chain-of-thought transparency, and OpenAI-compatible agentic tooling. Choose DeepSeek V3.2 if you need a larger 160K context window or lower output pricing.

Thanks to MoE architecture and MXFP4 quantization, it runs on a single 80GB GPU such as an NVIDIA H100 or AMD MI300X.

Yes. You can set low, medium, or high reasoning effort to trade latency against depth of thinking, and you get full access to the model's chain-of-thought for debugging.

Run GPT OSS 120B privately

No prompt logging. No data used for training.