For the first time in more than five years, OpenAI has released new open-weight language models designed to win back the hearts and workloads of developers around the globe. These latest models, gpt-oss-120b and gpt-oss-20b, are positioned as state-of-the-art solutions freely available for download, countering rapid gains made by OpenAI labs in China and rival US firms.
Developer-friendliness is front and center: the larger gpt-oss-120b can run on a single Nvidia GPU, while the lighter gpt-oss-20b suits laptops with just 16GB of memory. This represents a major improvement in accessibility compared to previous-generation open models and invites a broader range of experimentation.
More Than an Open Source Move
OpenAI, once a flagbearer for open-source innovation, has spent the last several years building a lucrative business around its closed-proprietary API and partnerships. The pressure to open up again has come both from policymakers in Washington and from surging open-source rivals like DeepSeek, Qwen, Moonshot AI, and Meta’s Llama series, which has recently lagged behind new entrants.
OpenAI CEO Sam Altman admits the company “has been on the wrong side of history” regarding open sourcing, and the gpt-oss series marks a direct response to changing tides and growing competitive urgency.
The move is also aligned with recent government calls for US AI companies to open technologies to boost global adoption of American-aligned AI.
Did you know?
OpenAI’s gpt-oss models are their first open-weight language models since GPT-2 was released in 2019, returning to a more open approach after years of favoring proprietary systems.
Performance, Hallucinations, and Guardrails
On industry-standard benchmarks, the models gpt-oss-120b and gpt-oss-20b achieved strong scores of 2622 and 2516 on the Codeforces coding test, respectively, outperforming Chinese open models such as DeepSeek R1, although they lag slightly behind OpenAI’s top-tier closed models.
The models also excel in general reasoning challenges like Humanity's Last Exam, outperforming most open competitors.
However, trade-offs remain clear. The open-weight models show a higher tendency to hallucinate: over 49% for gpt-oss-120b and 53% for gpt-oss-20b on OpenAI’s PersonQA benchmark, which is more than triple the rate of advanced proprietary models.
While mixture-of-experts design and reinforcement learning were used in training, OpenAI withheld the training data to address legal and copyright concerns, keeping some details under wraps.
ALSO READ | Anthropic Cuts OpenAI’s Claude API Access Citing Terms of Service Breach
Open, Yet Guarded
Developers are free to monetize and customize gpt-oss models thanks to the permissive Apache 2.0 license. However, OpenAI refrains from releasing training datasets, citing safety reviews and ongoing lawsuits related to data use.
The company insists these models were tested to prevent them from easily aiding in bioweapon development or cyberattacks and found only marginal risks compared to more advanced systems.
A Broader Battlefield
The gpt-oss models emerge as the US government accelerates support for open American AI, attempting to balance innovation with national values. The initial response from the developer community has been largely positive, with many welcoming the renewed openness and hinting at a wave of experimentation, product launches, and new research built atop OpenAI’s foundation.
Industry attention now shifts to upcoming releases from DeepSeek and Meta’s new superintelligence lab, as the open-weight arms race drives the next wave of global AI progress. OpenAI, for its part, is back in the open, ready to anchor the next generation of developer-driven breakthroughs.
Comments (0)
Please sign in to leave a comment