# GPT OSS 120B High ![Organization](https://img.shields.io/badge/Org-OpenAI-412991) ![Parameters](https://img.shields.io/badge/Params-120B-1f6feb) ![License](https://img.shields.io/badge/License-Apache%202.0-2da44e) ![Modality](https://img.shields.io/badge/Modality-Text-8957e5) ![Weights](https://img.shields.io/badge/Weights-Open-2da44e) > **Quick answer:** GPT OSS 120B High is OpenAI's first publicly released open-weight model — a 120-billion parameter model available for download and self-hosted deployment. "High" refers to the full-precision, highest-capability tier of the OSS 120B family (vs lighter quantised variants). It brings GPT-5-era training methodology to an open weights format, enabling research, fine-tuning, and private deployment. ## At a Glance **Where GPT OSS 120B High leads** - Open weights under Apache 2.0 — freely downloadable, self-hostable, and fine-tunable - 120B parameters trained with OpenAI's RLHF and alignment methodology — highest quality open model from OpenAI - No API rate limits, no per-token cost at inference — ideal for high-throughput or cost-sensitive private workloads - Suitable for domain fine-tuning (legal, medical, finance) without API data-handling restrictions **Where it lags** - Requires significant GPU infrastructure: at FP16, ~240GB VRAM (e.g., 3–4× A100 80GB or H100 80GB) - Capability ceiling below closed GPT-5.1 and GPT-5.2 — open weights trade some RLHF depth for accessibility - No multimodal capabilities in the OSS release (text only) **Best for:** enterprises needing on-premises deployment, researchers building fine-tuned domain models, and teams where data privacy restrictions preclude API-based inference. ## What GPT OSS 120B High Is GPT OSS 120B is OpenAI's strategic open-weight release, marking a reversal of the company's historically closed approach to model weights. The model was released in early 2026 as competitive pressure from open-weight leaders (Llama 4, Qwen 3, DeepSeek V3) intensified. "High" refers to the flagship full-precision variant within the OSS 120B family — OpenAI also offers quantised and smaller-parameter variants in the same release. The model was trained using a similar data pipeline and RLHF approach as the closed GPT-5 family, with the weights released under Apache 2.0 to maximise adoption. For Benchgen users, GPT OSS 120B High is interesting both as a self-hosted evaluation target and as a fine-tuning base for domain-specific agent evaluation. ## Specifications | Field | Value | |---|---| | Organization | OpenAI | | Parameters | 120 billion | | Architecture | Transformer decoder | | License | Apache 2.0 | | Release date | March 2026 | | Modality | Text | | Min VRAM (FP16) | ~240GB (3–4× A100/H100 80GB) | | HuggingFace | [openai/gpt-oss-120b](https://huggingface.co/openai/gpt-oss-120b) | ## Pricing GPT OSS 120B High weights are freely available under Apache 2.0. Self-hosted inference cost depends on hardware. OpenAI also provides hosted API access: | | Input (per 1M tokens) | Output (per 1M tokens) | |---|---|---| | OpenAI API (hosted) | ~$0.40 | ~$1.60 | ## GPT OSS 120B High vs Alternatives | Model | Params | License | Context | SWE-bench Verified | |---|---|---|---|---| | **GPT OSS 120B High** | 120B | Apache 2.0 | 128K | TBD | | [Llama 4 Maverick](/models/meta/llama-4-maverick) | 400B MoE | Llama 4 | 1M | ~50% est. | | [Qwen3 235B A22B](/models/alibaba/qwen3-235b-a22b) | 235B MoE | Apache 2.0 | 32K | ~55% est. | | [DeepSeek V3](/models/deepseek/deepseek-v3) | 685B MoE | MIT | 128K | ~42% | *(Scores are estimates pending Benchgen evaluations.)* ## Related Models & Benchmarks - Compare: [GPT-5](/models/openai/gpt-5) · [Llama 4 Maverick](/models/meta/llama-4-maverick) · [Qwen3 235B A22B](/models/alibaba/qwen3-235b-a22b) - Benchmarks: [SWE-bench Verified](/benchmarks/swe-bench-verified) · [MMLU-Pro](/benchmarks/mmlu-pro) ## Frequently Asked Questions

What is GPT OSS 120B High?

GPT OSS 120B High is OpenAI's first publicly released open-weight model — 120B parameters under Apache 2.0. "High" denotes the full-precision flagship tier of the OSS 120B family.

Can I fine-tune GPT OSS 120B High?

Yes — the Apache 2.0 license permits fine-tuning, derivative model releases, and commercial use without restriction.

What hardware is required?

At FP16, approximately 240GB VRAM — achievable with 3–4× A100 80GB or H100 80GB GPUs. Quantised variants (INT8, INT4) significantly reduce requirements.

--- *Specs based on OpenAI's GPT OSS 120B release. Last updated 2026-06-19.*

GPT OSS 120B High — Evaluations

Evaluations