AI provider OpenAI unveils adaptable AI models at no cost, detailing details such as cost, attributes, and further specifics
OpenAI Unveils Efficient AI Language Models, Paving the Way for Edge Deployment
OpenAI, the renowned artificial intelligence research laboratory, has announced the release of two new open-weight AI language models, gpt-oss-120b and gpt-oss-20b, on August 5. These models stand out for their use of a Mixture-of-Experts (MoE) architecture, which enables large total parameter counts while maintaining efficient inference.
The gpt-oss models were developed in partnership with Nvidia, Advanced Micro Devices, Cerebras, and Groq, allowing the models to be tested on various chips. Users can run these new models on PCs via LM Studio and Ollama.
OpenAI's latest release comes after tech giants like Meta, Microsoft-backed Mistral AI, and Chinese startup DeepSeek unveiled their open-weight models in the past few years. However, the key differences between OpenAI's gpt-oss models and those from other companies lie primarily in architecture, parameter efficiency, deployment scale, and performance characteristics.
OpenAI's gpt-oss Models
- Architecture: Both gpt-oss-120b and gpt-oss-20b use a MoE Transformer architecture, activating only a subset of parameters per token. This enables large total parameter counts but efficient inference.
- Parameters and Activation: gpt-oss-120b boasts 117 billion total parameters, activating 5.1 billion per token, while gpt-oss-20b has 21 billion total parameters, activating 3.6 billion per token.
- Deployment Efficiency: gpt-oss-20b can run on consumer-grade hardware with just 16 GB VRAM, suitable for edge devices or laptops. gpt-oss-120b requires a single 80 GB GPU (e.g., Nvidia H100) but is still efficient given its size.
- Performance: Both models match or surpass certain closed models from OpenAI’s own lineup across core reasoning benchmarks, coding, healthcare, math, and agentic tasks. They support long context windows (up to 128k tokens), tool use, chain-of-thought reasoning, and few-shot function calling.
- Licensing: Released under Apache 2.0 license, allowing wide reuse and integration.
- Quantization: They support 4-bit quantization enabling fast, memory-efficient inference.
Meta, Mistral AI, and DeepSeek Models
In comparison, Meta, Mistral AI, and DeepSeek generally release dense or mixture-of-experts models with varying sizes but tend not to emphasize the extreme parameter efficiency with selectively activated experts as aggressively as OpenAI's new gpt-oss MoE models. Model sizes from these firms vary widely; many focus on dense models optimized for either multi-GPU cloud inference or specialized hardware setups rather than desktop-grade hardware.
The new AI models, gpt-oss-120b and gpt-oss-20b, are free to download and users can customize them as needed. OpenAI's release of the new open-weight AI models was delayed multiple times, but last month, OpenAI CEO Sam Altman announced in an X post that additional safety tests and reviews of high-risk areas were required for the launch.
OpenAI President Greg Brockman stated that it is "exciting" to see the developing ecosystem and that the company is looking forward to contributing to it. These models have been designed to be lower-cost options and are freely available to download from the Hugging Face platform.
[1] OpenAI Blog: https://openai.com/blog/gpt-oss/ [2] Hugging Face: https://huggingface.co/blog/gpt-oss [3] Nvidia Blog: https://blogs.nvidia.com/blog/2023/08/22/openai-releases-gpt-oss-models-nvidia-platforms/ [4] AMD Blog: https://www.amd.com/en/blog/openai-releases-gpt-oss-models-on-amd-eypc-instances [5] Cerebras Blog: https://www.cerebras.net/blog/openai-releases-gpt-oss-models-on-cerebras-wafer-scale-engine-2
- OpenAI's recent innovation in artificial intelligence, the gpt-oss models, have been developed with a Mixture-of-Experts (MoE) architecture, which allows for efficient utilization of technology in language models, making them suitable for edge devices, such as laptops and even cricket score prediction systems that leverage AI.
- As the field of technology evolves, companies like OpenAI are pushing the boundaries of artificial intelligence with latest advancements like MoE Transformers and 4-bit quantization, revolutionizing the way we interact with technology, whether it's playing a game of cricket or engaging in complex tasks like coding and healthcare.
