OpenAI GPT OSS 120B
OpenAI's largest open-weight reasoning model — 117B MoE parameters, agentic tool use, and full chain-of-thought under Apache 2.0.
Get API key- Provider
- OpenAI
- Price
- $0.07 in · $0.30 out / 1M
- Context window
- 128K tokens
- Released
- August 5, 2025
- License
- Apache 2.0
What is OpenAI GPT OSS 120B?
OpenAI GPT OSS 120B is an open-weight, text-only reasoning model released in August 2025. It features a 117B-parameter mixture-of-experts architecture with 5.1B active parameters, native function calling, web search, and adjustable reasoning effort, distributed under the Apache 2.0 license for production use.
Use it privately on Venice
On Venice, GPT OSS 120B runs under a private, zero-retention tier — your prompts are never stored, profiled, or used for training. You can tap its native tool use, reasoning, and web search capabilities without Big-Tech surveillance, paying only for the tokens you consume.
What can it do?
- •Open weights under the permissive Apache 2.0 license: freely customizable, fine-tunable, and deployable without copyleft or patent risk.
- •Strong reasoning with configurable effort (low, medium, high) and full chain-of-thought visibility for debugging and trust.
- •Native agentic capabilities including function calling, web search, Python code execution, and Structured Outputs.
- •Efficient inference — MXFP4-quantized MoE weights fit on a single 80GB GPU such as an NVIDIA H100 or AMD MI300X.
- •Competitive benchmark performance — near-parity with OpenAI o4-mini on core reasoning tasks and strong results on agentic evaluations like Tau-Bench.
- •Text-only — no vision, image, or audio input support.
- •Requires the harmony response format; using standard chat templates produces incorrect or degraded outputs.
- •Safety responsibility shifts to the deployer — open weights can be fine-tuned to bypass refusals, so downstream systems need their own guardrails.
- •128K context window is smaller than million-token or 256K rivals available on Venice.
- •Not uncensored by default — it follows OpenAI safety policies out of the box, though the open-weight release makes downstream modification possible.
OpenAI GPT OSS 120B capabilities
- 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
How to use it via API
Venice exposes an OpenAI-compatible API. Swap your base URL and call openai-gpt-oss-120b.
curl https://api.venice.ai/api/v1/chat/completions \
-H "Authorization: Bearer $VENICE_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "openai-gpt-oss-120b",
"messages": [{ "role": "user", "content": "Explain quantum tunneling simply." }]
}'Specifications
Pricing
Billed per token on Venice: $0.07 per 1M input tokens and $0.30 per 1M output tokens.
New Venice accounts include a free daily allowance and 500 welcome credits — no credit card required.
OpenAI GPT OSS 120B vs alternatives
| Model | Context window | Strongest at | Open weights | Price (Venice) |
|---|---|---|---|---|
| OpenAI GPT OSS 120B | 128K tokens | Open reasoning & tool use | Yes | $0.07 in · $0.30 out / 1M |
| DeepSeek V3.2 | 160K tokens | Efficient MoE reasoning | Yes | $0.33 in · $0.48 out / 1M |
| Claude Opus 4.8 | 1M tokens | Closed frontier reasoning | No | $6 in · $30 out / 1M |
| Kimi K2.6 | 256K tokens | Long-context open weights | Yes | $0.75 in · $3.50 out / 1M |
OpenAI's largest open-weight model: efficient MoE inference with full CoT and agentic capabilities.
What is it good for?
- •Agentic workflows that combine reasoning with tool use, web search, and code execution.
- •Private or on-premise deployment on a single H100 GPU for enterprises requiring data sovereignty.
- •Fine-tuning and customization for domain-specific reasoning, instruction following, or structured output tasks.
- •Debugging complex logic via full chain-of-thought access without exposing it to end users.
- •Low-cost, high-quality text generation via the Venice API without subscription lock-in.
Prompting tips
- •Always use the harmony response format; standard chat templates will cause the model to malfunction.
- •Set reasoning_effort to low for simple queries to reduce latency and cost, and high for math or logic problems.
- •Use native function calling and web search instead of prompting for raw facts — the model is optimized for agentic tool use.
- •Inspect the chain-of-thought for debugging and verification, but filter it before showing outputs to end users.
Version history
Smaller 21B-parameter variant for edge and local deployment within 16GB memory.
CurrentCurrent 117B-parameter production model for high-reasoning agentic workflows.
Frequently asked questions
OpenAI GPT OSS 120B is an open-weight, text-only reasoning model released on August 5, 2025. It features a 117B-parameter mixture-of-experts architecture with 5.1B active parameters, native function calling, web search, and adjustable reasoning effort, distributed under the Apache 2.0 license.
Venice bills per token with zero retention: $0.07 per 1M input tokens and $0.30 per 1M output tokens. There is no subscription required; you pay only for what you use.
Yes. The model is released under the Apache 2.0 license with open weights hosted on Hugging Face. You can download, fine-tune, and self-host it, though it requires the harmony response format to work correctly.
The weights are free to download and use under the Apache 2.0 license. On Venice, inference is billed per token; new accounts may receive free credits to try the model, but sustained use requires payment.
Yes. It supports native function calling, web search, Python code execution, and Structured Outputs, making it well-suited for agentic workflows.
Claude Opus 4.8 offers a 1M-token context window and closed-frontier performance at a premium price. GPT OSS 120B is far cheaper, runs on a single H100 GPU, and gives you open weights with full chain-of-thought transparency — choose it if you need sovereignty, local deployment, or transparent reasoning.
Yes. With MXFP4 quantization, the model fits on a single 80GB GPU such as an NVIDIA H100 or AMD MI300X. You can download the weights from Hugging Face and run inference locally or on private cloud infrastructure.
Harmony is the specific chat format the GPT OSS models were trained on. They must be prompted with this format; using standard chat templates will cause incorrect or degraded outputs.
Related models
Run OpenAI GPT OSS 120B privately.
No prompt logging. No data used for training. Free to start — no credit card.
