furiosa-ai/gpt-oss-20b

TEXT GENERATIONPricing:Input $0.3 / Output $1.2Concurrent Unit Cost:1Model Size:20BQuant:FP8Context Size:32kPublished:Jun 9, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

furiosa-ai/gpt-oss-20b is a 20 billion parameter Mixture-of-Experts (MoE) transformer model developed by OpenAI, optimized for lower-latency and on-device reasoning tasks. It utilizes the harmony response format to separate chain-of-thought reasoning from the final answer and natively supports tool calling with configurable reasoning effort. This model is specifically packaged with a Furiosa Executable Bundle (FXB) for efficient deployment on FuriosaAI RNGD hardware, with expert weights quantized to MXFP4 for performance.

Loading preview...

Overview

furiosa-ai/gpt-oss-20b is a 20 billion parameter Mixture-of-Experts (MoE) transformer model, originally developed by OpenAI, and optimized for lower-latency and on-device use cases. This version is specifically packaged with a Furiosa Executable Bundle (FXB) for deployment on FuriosaAI RNGD hardware, leveraging the Furiosa-LLM inference engine. It maintains the core features of the upstream openai/gpt-oss-20b model, including its unique reasoning capabilities and tool-calling support.

Key Capabilities

  • Advanced Reasoning: Employs a chain-of-thought approach, separating reasoning steps from the final answer using the harmony response format. Reasoning depth is controllable via reasoning_effort (low, medium, high).
  • Tool Calling: Natively supports tool (function) calling, compatible with the OpenAI tool-call parser.
  • Optimized for FuriosaAI: Packaged for efficient execution on FuriosaAI RNGD, utilizing MXFP4 quantization for MoE expert weights while maintaining higher precision for other components.
  • OpenAI API Compatibility: The Furiosa-LLM server exposes an OpenAI-compatible API for easy integration.

Good for

  • On-device AI applications: Designed for scenarios requiring lower latency and efficient execution on edge or specialized hardware.
  • Reasoning-intensive tasks: Ideal for applications where explicit chain-of-thought and controllable reasoning depth are beneficial.
  • Tool-augmented language models: Suitable for use cases requiring robust function calling and integration with external tools.
  • Deployment on FuriosaAI hardware: Optimized for performance and efficiency when run on FuriosaAI RNGD cards, utilizing a tensor-parallel size of 8 PEs per card.

Finetunes
1 models