Artificial Intelligence & Machine Learning / Open-source AI models

Open-source AI models often require more computing power because they may not be as optimized for efficiency as closed-source models. This leads to higher resource utilization for equivalent tasks, increasing energy consumption and operational costs over time.

'Compute' refers to the hardware resources such as CPUs, GPUs, or TPUs used to perform the numerical calculations necessary for training and running AI models. It is measured in FLOPS (floating-point operations per second). The amount of compute determines how quickly and effectively a model can learn from data and perform tasks, impacting both performance and energy consumption.

The gpt-oss-120B and gpt-oss-20B are OpenAI's new open-weight language models released under the Apache 2.0 license. The 120B model has 117 billion parameters and activates 5.1 billion parameters per token using a mixture-of-experts architecture, while the 20B model has 21 billion parameters and activates 3.6 billion parameters per token. These models are designed for efficient deployment, with the 20B model able to run on consumer-grade hardware with just 16 GB of RAM, making it suitable for edge and on-device use. They demonstrate strong reasoning, tool use, and structured output capabilities, approaching or surpassing the performance of proprietary models like OpenAI's o4-mini and o3-mini on various benchmarks.

'Open-weight' means that the full model weights are publicly available under a permissive Apache 2.0 license, allowing developers to run, modify, and deploy the models without restrictions typical of proprietary models. This is significant because it enables local and edge deployment, reducing reliance on cloud infrastructure. The availability of these models on AWS for the first time means that users can access and run these open models directly on Amazon Web Services, facilitating scalable cloud-based use and integration into existing AWS workflows.

OpenAI’s new models, gpt-oss-120b and gpt-oss-20b, are released with open weights under the permissive Apache 2.0 license, meaning anyone can access, run, and modify the models locally without relying on OpenAI’s API. This is significant because previous models like GPT-3.5 and GPT-4 were closed-source and API-only, limiting control over latency, cost, and privacy. Open-source availability enables developers and companies to deploy powerful AI on consumer hardware, fostering innovation and reducing infrastructure costs.

The gpt-oss-120b model matches the performance of OpenAI’s o4-mini model on reasoning benchmarks while requiring significantly less hardware (a single 80GB GPU). The smaller gpt-oss-20b model performs similarly to o3-mini and can run on consumer devices with 16GB GPU. These models excel in advanced reasoning, tool use (such as web search and code execution), and support chain-of-thought reasoning and structured outputs, outperforming other open-source models of similar size.

DeepSeek AI uses a Mixture-of-Experts (MoE) architecture, which consists of multiple specialized sub-models or 'experts' trained for specific tasks. When processing a query, only the most relevant experts are activated instead of the entire model. This selective activation reduces computational load, increases efficiency, and improves accuracy by leveraging specialized expertise for different types of data or tasks.

Open-source software development allows AI models like DeepSeek to be more transparent, accessible, and collaboratively improved by the tech community. This approach accelerates innovation, enables diverse contributions, and helps balance the risks and rewards associated with AI development. The rise of generative AI has intensified debates about open-source's role, highlighting its growing importance in shaping the future of AI technology.

OpenAI has released two new open-weight AI reasoning models called gpt-oss-20B and gpt-oss-120B. These models are open-source, allowing enterprises to customize them by training on specific corporate data, which enhances tailored AI solutions. Unlike OpenAI's recent proprietary models, these open models can be run locally on hardware ranging from consumer laptops to single Nvidia GPUs, and they support advanced reasoning and tool use. This marks OpenAI's first open model release since GPT-2, over five years ago.

CIOs need to assess risks related to data security, privacy, and governance when deploying OpenAI's open models. Although these models enable customization on corporate data, enterprises must ensure proper usage governance, including logging, guardrails, and personally identifiable information (PII) detection, to prevent data leaks or misuse. Additionally, integrating open models with proprietary cloud AI services may introduce complexity and require careful risk management to balance innovation with security.

OpenAI's open-weight models, such as gpt-oss-20B and gpt-oss-120B, are AI models released with open weights under a permissive Apache 2.0 license, allowing anyone to run and customize them on their own infrastructure. These models support advanced reasoning and tool use, enabling complex multi-step tasks and domain-specific AI applications. Their open availability marks a major step in democratizing AI access, allowing developers, enterprises, nonprofits, and governments worldwide to build and innovate without relying solely on proprietary models. This openness also lowers barriers for emerging markets and smaller organizations, fostering broader AI-driven economic growth and innovation globally[1][2][3][4].

While OpenAI's open-weight models demonstrate impressive reasoning capabilities and flexibility, they come with a tradeoff in terms of hallucinations—instances where the AI generates plausible but incorrect or fabricated information. Additionally, these models may require more customization and integration effort compared to closed-source models that offer more turnkey solutions. However, OpenAI mitigates some limitations by enabling these open models to connect with more capable closed models in the cloud for tasks they cannot perform alone, such as image processing. This hybrid approach balances openness with performance and reliability[2][3].

The Apache 2.0 license is a permissive open-source license that allows users to freely use, modify, distribute, and sublicense the AI models, including for commercial purposes. Users must include the original copyright notice and license text, and disclose significant changes made to the code. This license encourages innovation and broad usage without API fees, enabling deployment on consumer hardware or cloud platforms.

Under OpenAI's terms, users retain ownership of both their input data and the output generated by the models. The output can be licensed under permissive licenses like Apache 2.0 or MIT by the user, allowing them to share, modify, or use the data freely, including for training other models. Restrictions apply only to the original user, not to the data itself, promoting openness and reuse within the community.

'Open-weight' means that the full model weights (parameters) are publicly released and available for free download, allowing developers and researchers to run, fine-tune, and customize the models without restrictions. This contrasts with proprietary models where weights are kept private and only accessible via API. OpenAI’s gpt-oss-120b and gpt-oss-20b models are open-weight, licensed under Apache 2.0, enabling experimentation and commercial use without copyleft or patent risks.

The gpt-oss-20b model is smaller (21 billion parameters) and optimized to run with lower latency and within limited memory (16GB), making it suitable for local use on consumer hardware like Apple Silicon Macs. Its native MXFP4 quantization and efficient architecture allow it to operate without requiring large, specialized GPUs, unlike the larger gpt-oss-120b which needs more powerful hardware such as an 80GB GPU.

The gpt-oss-120b model has 117 billion total parameters and activates about 5.1 billion parameters per token, delivering stronger reasoning and better performance on complex tasks. It requires a high-end GPU like the NVIDIA H100 or a multi-GPU setup. The gpt-oss-20b model has 21 billion total parameters, activates about 3.6 billion parameters per token, and is optimized for speed and accessibility, fitting on a single 16GB GPU, making it suitable for on-device or low-cost server inference. Both models use Mixture-of-Experts (MoE) architecture with 4-bit quantization for efficient inference but differ in scale and hardware requirements.

'Open-weight' means that the full model weights of gpt-oss-120b and gpt-oss-20b are publicly released under a permissive Apache 2.0 license, allowing developers and researchers to freely access, run, modify, and fine-tune these models. This openness promotes transparency, wider accessibility, and innovation in AI development by enabling experimentation without the restrictions of proprietary models. It also allows deployment on consumer hardware and supports diverse applications such as tool use, chain-of-thought reasoning, and agentic tasks.

Open-weight AI models are artificial intelligence models whose internal parameters (weights) are made publicly accessible, allowing researchers and developers to study, modify, and build upon them. OpenAI's release of two open-access models follows the success of China's DeepSeek, which gained global attention for its innovative AI software. This move by OpenAI represents a significant advancement in AI, promoting transparency and wider collaboration in replicating human reasoning capabilities.

OpenAI's new open-weight models are designed to replicate human reasoning and follow DeepSeek's success in this area. DeepSeek's R1 model slightly outperforms OpenAI's o1 model in some reasoning benchmarks, such as mathematical reasoning (DeepSeek-R1 scored 97.3% vs. OpenAI o1's 96.4%) and operates at a lower cost. However, OpenAI's models maintain strong coding capabilities and have recently released frontier models like o3 that surpass DeepSeek's R1 in overall performance. This competitive dynamic has driven OpenAI to release open-weight models to foster innovation and accessibility.

'Open-weight' means that the full model parameters (weights) are publicly available for download and use, allowing researchers and developers to run, modify, and build upon the models without restrictions. This contrasts with proprietary models where weights are kept private and only accessible via API.

The new open-weight models, gpt-oss-120b and gpt-oss-20b, focus on text-based reasoning and generation and do not support multi-modal inputs like images or audio. This limitation likely reflects a trade-off to prioritize accessibility and flexibility for users while maintaining manageable model complexity and hardware requirements.

'Open-weight' means that the full model parameters (weights) are publicly available for download, modification, and deployment without restrictive licensing. This allows developers and researchers to customize, fine-tune, and run the models on their own hardware, unlike closed models that only provide API access.

OpenAI is releasing open-weight models to reestablish its presence in the open model landscape amid rising competition from Chinese AI models and Meta's shift towards closed releases. This move also reflects a balance between openness and safety, as OpenAI has introduced new safety protocols to mitigate risks associated with open-source models.

Iterated Distillation and Amplification (IDA) is a training technique where the AI model internalizes the reasoning process through iterative policy improvement rather than relying on longer search times during inference. This method enables Deep Cogito v2 models to learn more efficient and accurate reasoning skills, improving performance on complex tasks such as math and language benchmarks while remaining cost-effective.

Deep Cogito v2 models are called hybrid reasoning models because they can toggle between two modes: a fast, direct-response mode for simple queries and a slower, step-by-step reasoning mode for complex problems. This hybrid approach allows the models to efficiently handle a wide range of tasks by balancing speed and depth of reasoning, outperforming other open-source models of similar size.