In brief: OpenAI announces the release of OSS 20B and OSS 120B
OpenAI is opening access to OSS 20B and OSS 120B — the first open-weight models since GPT-2. These new language models can be used locally on regular laptops as well as on powerful workstations with NVIDIA RTX. OSS models offer scalability and transparency, since the full set of parameters is available for customization, testing, and integration into business processes.
In GPTunneL both models are available to all users without the need to configure the model on their own devices. Just pick the model, enter your prompt, and use it.
The release of OSS 20B and OSS 120B under the Apache 2.0 license strengthens the position of open-source artificial intelligence as an industry standard and lowers the barrier to entry for developing large LLM solutions. This event is shaping new trends in open AI and driving the growth of open-source communities.
Why the release matters for the industry, developers, and companies
OSS models allow neural networks to run locally, without sending business data to the cloud and without extra infrastructure costs. This also gives users in regions with restrictions a chance to access the latest technologies.
For developers, this is a launchpad for building AI solutions that can be quickly adapted to their industry's specifics. Companies get tools for analyzing documentation, automating support, and custom text generation with minimal risk of leaks.
- OpenAI OSS removes dependence on commercial services and gives full control over how the model is used;
- Large language models accelerate the adoption of AI solutions even in regulated fields — finance, medicine, engineering;
- Testing new models and fine-tuning are available without closed licenses — especially relevant for research groups;
- The release of OSS 20B and OSS 120B has become a notable factor influencing the market and the open-source trend for business.
OpenAI OSS, DeepSeek, Mistral, and Meta now compete on equal footing for the attention of enthusiasts, specialists, and enterprise IT teams.
Key features of OSS 20B and OSS 120B
OSS 20B is a "lightweight" model with 21 billion parameters, designed to run on hardware with 16 GB of RAM: from workstations to Copilot+ AI laptops.
The powerful OSS 120B contains 120 billion parameters, but can run on a single 80 GB A100 GPU and processes tokens efficiently thanks to its Mixture-of-Experts (MoE) architecture.
This technology allows only a few "experts" to be activated per token instead of the entire network, which makes inference cheaper and speeds up analysis, generation, and coding tasks.
OpenAI notes that OSS 120B's performance on complex medical and expert-level tasks surpasses o4-mini. The context window in GPT-OSS 20B/GPT-OSS 120B, for example through the GPTunneL service, reaches 131,000 tokens — enough to load technical documentation, analytical data, or a long support conversation without losing context.
- OSS 20B: 21 billion parameters, 16 GB RAM, fast launch on average PCs;
- OSS 120B: 120 billion parameters, 80 GB RAM, processing speed on par with the best closed-source alternatives;
- Context: up to 131,000 tokens, suitable for working with large texts and code;
- MoE architecture minimizes hardware requirements while preserving output quality;
New capabilities of OSS models
OpenAI's open models bring new use cases for business, R&D, and enthusiasts. You can now build chatbots that analyze inquiries using your own data, automate report generation and research writing without sharing information with third parties.
For example, a law firm implements OSS 20B for automatic contract analysis, while a financial company customizes OSS 120B to detect anomalies in large transaction datasets. Within GPTunneL, users run both models with full functionality, interacting with them through chat, a Telegram bot, or via API integration.
Here are the main use cases:
- Automating support and processing routine requests;
- Flexible code generation with explanations for development teams;
- Recommendation systems with long "memory" and personalization (for example, curating best practices for HR at an IT company);
- Rapid testing and benchmarking of new models in a corporate sandbox — without license restrictions.
Using OSS makes open-source adoption easier for business — from regional banks to small digital teams that need to quickly adapt neural networks for document analysis, text intelligence, or auditing new markets. It's worth noting an important feature of OSS models: transparency of the reasoning chain, which makes it possible to understand exactly how the final output was formed and speeds up content validation.
Access and terms of use
OSS 20B and OSS 120B can be obtained through Hugging Face, Azure, and AWS — the models are distributed under the OSS license, which permits free fine-tuning, integration, and commercial distribution with attribution. To get started, it's usually enough to register on the platform and confirm agreement with terms that prohibit generating harmful content (CBRN, prompt injections, etc.).
In GPTunneL, users get validated instances of the models with quick setup, testing, a convenient API, and support for custom assistants. OSS AI models are ideal for local experiments and integration into products where data is critical for the company.
Summary and outlook
The launch of OSS 20B and OSS 120B is strengthening the influence of open-source solutions on the market:
- Major players can no longer ignore the demand for transparency, fast adoption, and long-term customization.
- OSS models are reshaping the industry: they make it possible to build trustworthy solutions within an organization, raise the level of AI literacy, and accelerate the integration of neural networks into complex business processes.
Open AI trends are shifting the focus toward customized solutions and reduced dependence on the cloud. In the coming years we expect even greater growth of OSS models — competitors are already responding to the release of OSS 20B and OSS 120B with their own lineups.
For example, Mark Zuckerberg, head of Meta, recently shared the company's vision for the development of AI: they intend to build artificial superintelligence for all users, combining open and closed models.
Open models are becoming the benchmark for developing scalable AI solutions: their market impact keeps growing, and communities are expanding the testing and task ecosystem. The time has come when adopting neural networks is not about chasing hype, but a strategic decision for effective, manageable, and intelligent systems. The GPTunneL team constantly monitors AI news and adds new solutions for all users. Try OSS 20B and OSS 120B in GPTunneL!