OpenAI gpt-oss Release on Ollama

Ollama and OpenAI have partnered to bring the gpt-oss open weight models to the Ollama platform, enabling local execution of state-of-the-art reasoning and agentic models. This release introduces two model sizes—20B and 120B parameters—designed for versatile developer use cases and high-reasoning tasks.

Core Capabilities and Features

The gpt-oss models provide native support for agentic workflows and transparent reasoning processes. Key features include:

  • Agentic Capabilities: The models natively support function calling, python tool calls, and structured outputs. Ollama provides an optional built-in web search to augment the models with real-time information.
  • Full Chain-of-Thought: Users have complete access to the model's internal reasoning process, which is intended to increase output trust and simplify debugging.
  • Configurable Reasoning Effort: The reasoning effort can be adjusted between low, medium, and high settings to balance latency and performance based on the specific use case.
  • Customization and Licensing: The models are fully fine-tunable via parameter fine-tuning and are released under the permissive Apache 2.0 license, allowing for commercial deployment and experimentation without copyleft restrictions.

Model Variants and Hardware Requirements

Two versions of gpt-oss are available to suit different performance and hardware constraints:

  • gpt-oss-20b: A 20 billion parameter model optimized for lower latency and specialized local use cases.
  • gpt-oss-120b: A 120 billion parameter model designed for general-purpose, high-reasoning production tasks.

Technical Implementation: MXFP4 Quantization

To reduce memory footprints, OpenAI utilizes the MXFP4 quantization format for the mixture-of-experts (MoE) weights, which account for over 90% of the total parameter count. In this format, weights are quantized to 4.25 bits per parameter.

This quantization enables the following hardware accessibility:

  • The 20B model can run on systems with as little as 16GB of memory.
  • The 120B model can fit on a single 80GB GPU.

Ollama has developed new kernels for its engine to support the MXFP4 format natively, avoiding the need for additional conversions. Ollama collaborated with OpenAI to benchmark these implementations against reference versions to ensure quality parity.

Hardware Acceleration with NVIDIA

Ollama and NVIDIA have collaborated to optimize gpt-oss performance specifically for NVIDIA GeForce RTX and RTX PRO GPUs. This partnership ensures that users with RTX-powered PCs can accurately leverage the full capabilities of the gpt-oss models.

Deployment

The models can be deployed via the latest version of the Ollama application or through the terminal using the following commands:

ollama run gpt-oss:20b

ollama run gpt-oss:120b

Sources

Related

  • Project
    openai/gpt-oss

    OpenAI’s gpt‑oss repo delivers reference code and tooling for the open‑weight 120 B and 20 B LLMs, including a special Harmony chat format, MXFP4 quantization, and back‑ends for PyTorch, Triton, Metal, vLLM, Ollama, and more. Licensed Apache‑2.0, it’s aimed at researchers and developers who want to run, fine‑tune, or build tool‑augmented agents with these models.

  • Dispatch
    OpenAI gpt-oss Release Notes

    OpenAI has released gpt-oss-120b and gpt-oss-20b, two open-weight reasoning models licensed under Apache 2.0 that deliver frontier-level reasoning and tool-use capabilities optimized for consumer hardware.

  • Dispatch
    OpenAI Codex with Ollama Integration

    Ollama now enables the OpenAI Codex CLI to use open-weight models, allowing for local code execution and modification using models like gpt-oss:20b and gpt-oss:120b.

  • Dispatch
    OpenAI gpt-oss-120b and gpt-oss-20b Release

    OpenAI has released gpt-oss-120b and gpt-oss-20b, two open-weight reasoning models under the Apache 2.0 license designed for agentic workflows and customizable reasoning effort.

  • Dispatch
    OpenAI gpt-oss-safeguard models released via Ollama

    OpenAI and Ollama released the open‑source gpt-oss-safeguard safety‑reasoning models (20B and 120B) under Apache 2.0, enabling organizations to run policy‑driven content classification with transparent reasoning.