OpenAI drops two open models ahead of GPT-5 launch: GPT OSS 120B and 20B explained in 5 points
OpenAI says the models are tailored for agent-style tasks and include support for advanced reasoning workflows. While these open models cannot directly process complex data like images, they can route queries to OpenAI's more powerful closed models via cloud APIs, effectively acting as intelligent intermediaries.The models rely on a Mixture-of-Experts (MoE) architecture, which allows them to activate only a small subset of parameters per token, around 5.1 billion for the 120b model, ensuring greater efficiency and responsiveness. A post-training process involving high-compute reinforcement learning further enhances their reasoning abilities, aligning them closely with OpenAI's o-series of frontier models.-- OpenAI claims its open models set a new benchmark in the open-weight category. On Codeforces, a widely used programming benchmark, GPT-OSS-120b scored 2622, and the smaller 20b model scored 2516 — both outperforming DeepSeek's R1 but still behind OpenAI's o3 and o4-mini.-- However, hallucination remains a significant issue, as reported by TechCrunch. On OpenAI's PersonQA benchmark, a test of factual accuracy regarding people, GPT-OSS-120b hallucinated 49 per cent of the time, while the 20b version did so in 53 per cent of cases. That's substantially worse than the 16 per cent hallucination rate observed in OpenAI's older o1 model, and even higher than the 36 per cent rate for o4 mini.OpenAI attributes this to the reduced parameter activation and narrower 'world knowledge' of smaller models, an expected trade-off when moving away from large frontier systems.-- In a white paper accompanying the launch, OpenAI addressed concerns over potential misuse. The company says both internal and third-party evaluations were conducted to assess the risk of the models being repurposed for cybercrime or biochemical threats. While GPT-OSS could marginally increase biological knowledge in bad actors, OpenAI concluded it does not meet its 'high capability' danger threshold, even after fine-tuning.Unlike some fully open-source labs such as AI2, OpenAI has chosen not to release the training datasets used to build GPT-OSS, a likely response to ongoing lawsuits alleging copyright misuse in AI training.advertisementStill, the Apache 2.0 licence grants developers wide-ranging freedom, including commercial use, without any obligation to pay OpenAI or seek further permissions. This could significantly boost adoption, particularly among startups and enterprises aiming to integrate capable AI models without high licensing costs.-- After years of a tightly guarded, proprietary approach, OpenAI has taken a decisive turn by re-entering the open source arena. The move is likely an attempt to reassert leadership in a space increasingly dominated by Chinese players such as DeepSeek, Moonshot AI, and Alibaba's Qwen, all of whom have recently made major strides in open model development. While Meta's Llama once led this frontier, its influence has waned over the past year.With the release of GPT-OSS, OpenAI is making a calculated bid to re-engage developers and policymakers, particularly the Trump administration, which recently urged US firms to open source more AI tech to advance democratic values in global AI adoption. CEO Sam Altman admitted earlier this year that OpenAI may have been 'on the wrong side of history' when it came to transparency and openness.- Ends
Try Our AI Features
Explore what Daily8 AI can do for you:
Comments
No comments yet...
Related Articles
Time of India
an hour ago
- Time of India
Free GPT-5 access heats up AI rivalry: Who wins, who loses?
The past fortnight has been incredible in terms of the buzzing AI space, particularly the open source AI domain. First with the coming of Perplexity Comet, and now with the release of GPT 5, the sector is literally taking neckbreak competition to another level. OpenAI has unveiled GPT-5, its most sophisticated artificial intelligence model yet, and the best part is that it's now available for free to all ChatGPT users—albeit with some usage the first time, free-tier users can access a reasoning-capable AI model—GPT-5—unlocking advanced capabilities such as multi-step reasoning, better understanding of complex prompts, and significantly improved accuracy. When users hit their usage cap, the system seamlessly pivots to a lighter but still capable version called GPT-5 Mini. Plus-tier subscribers can enjoy elevated limits, while Pro and Team users gain unlimited access, including access to GPT-5 Pro, a version tailored for deep, intricate reasoning tasks. GenAI race gets hotter "This is a very happening space. Sam Altman sees the GPT-5 as a step toward ethical AI but warns against over-reliance. Elon Musk, on the other hand, warns of OpenAI's growing influence, citing geopolitical risks. There are other leaders who call it solid but overhyped, exposing LLM limits and pushing hybrid AI approaches," says Gopi Thangavel, Group CIO at Larsen & Toubro, commenting on the development. "This can be the way to crush the competition -- a strategy used by most. For example, Perplexity is free for India. Moreover, OpenAI has removed other models as they see GPT-5 as more stable and advanced compared to previous models. If you see, they timed it well -- giving previous lighter models as free as OSS 20b and 120b and then launched GPT-5. Besides, GPT 5 is much better than previous models in coding," says Gaurav Rawat, an AI leader with a leading financial sector company and who has experience across the auto and pharma sectors as well. Designed on the GPT 4 model, GPT-5 is able to offer better detailed and thoughtful answers by analysing intricate tasks such as science questions, coding, information synthesis, or financial analysis. It's also able to return more accurate as well as faster responses as compared to its past models. The new model has the potential to work on various data forms, including images and text, all at the same time. This primarily translates into the fact that somebody can ask the model a question, put up a picture, or even share a document. The GPT 5 can comprehend as well as respond without the need of separate tools. Such a capability wasn't there before. Free GPT5 sustainable? "This development is reflective of the competition in this space, which is huge. Other than a few, LLMs are struggling to make money. They spend a lot developing them but are unable to get it back," says Avik Sarkar, Senior Research Fellow & Visiting Faculty at Indian School of Business. "Nothing is ever truly free — when it's free, you are the product. OpenAI's decision to make GPT‑5 accessible to everyone is a smart play: offer a preview, gauge user engagement, shape future pricing, grow the customer base — and at the same time, turn up the competitive heat. It's the same psychology as letting readers preview a few Kindle pages before buying the book — just at an AI scale. This move will fortify Microsoft's ecosystem and influence. But make no mistake — GPT‑5's capabilities will reshape the competitive landscape. The real question: Which rival will be the first to reveal what they've been hiding under the hood?" comments Rajendra Deshpande, former CIO at Intelenet Global Services.
Mint
5 hours ago
- Mint
Income Tax: Taxpayers need to use right ITR form to file returns. Here's how to choose it
Income Tax: Fewer than 30 days are remaining before the September 15 deadline to file the income tax return (ITR) for FY 2024-25. It is crucial for taxpayers to use the correct income tax return form to avoid rejection at the time of filing it. Notably, the income tax department has released the excel utilities of ITR 1 to ITR 6 for filing tax returns for FY 2024-25. It is important to note that the I-T Department on 16 August released the excel utilities for ITR-3 and ITR-4 for filing updated returns for FY 2021-22 and FY 2022-23. Prior to this, the excel utilties for filing updated returns were released for ITR-1 and ITR-2 on 29 July. Here, we list the key features of the most common income tax return forms — ITR-1 to ITR-4. ITR-1 (Sahaj): This income tax form is meant for resident individuals who have a total income of up to ₹ 50 lakh. Different sources of income entail salary/pension, income from one house property (barring the cases where loss is carried forward) and income from other sources (such as interest from savings accounts and FDs ), along with agricultural income (for a maximum of up to ₹ 5,000). It may be noted that you are not eligible if one of these is true: you are a director of a company and invested in unlisted equity shares during previous year, you have income from business or profession, or you are a resident owning foreign assets or foreign income or more capital gains than allowed threshold with long- term capital gains under section 112A which is more than ₹ 1.25 lakh or having carried forward/ brought forward losses. ITR-2: ITR-2 can be filed by individuals or Hindu Undivided Families (HUFs) who are not eligible to file the Sahaj form. Taxpayers who do not have income from profits and gains of business or profession, and also do not have income from profits and gains of business or profession in the nature of interest, salary, bonus and commission or remuneration from a partnership firm can use this form. This income tax return form can also be filed by those taxpayers who want the income of another person, such as their spouse or a minor child, to be clubbed with their own income, provided the income that needs to be clubbed falls in one of the specified categories. Date Tax return form released 14 Aug 2025 ITR-6 excel utility released 8 Aug 2025 ITR-5 excel utility released 30 July 2025 ITR-3 offline and online utility 17 July 2025 ITR-2 online utility released 11 July 2025 ITR-2 and ITR-3 excel utilities released 4 June 2025 ITR-1 and ITR-4 online utilities released 29 May 2025 ITR-1 and ITR-4 excel utilities released (Source: Income Tax Department) ITR-3: This income tax return form is meant for individuals and HUFs engaged in business or professions requiring the maintenance of elaborate books of accounts. This category also entails professionals, including doctors, lawyers, and chartered accountants, whose income is calculated based on actual profits. ITR-4 (Sugam): This income tax form can be filed by a resident individual/ HUF/ Firm (other than LLP) who has income up to ₹ 50 lakh during a year, income from business and profession computed on a presumptive basis under sections 44AD, 44ADA or 44AE, income from salary/pension, one-house property, agricultural income (up to ₹ 5,000) and other sources. For more details on income under the presumptive tax scheme, visit this link. For all personal finance updates, visit here
Time of India
6 hours ago
- Time of India
OpenAI-rival Anthropic sets limits on how investors can participate in upcoming $5 billion fundraise
Claude-maker Anthropic has told investors that the AI company does not want money coming through special purpose vehicles (SPVs) in its latest fundraising. Citing two people familiar with the matter, a Business Insider report said that one of Anthropic's largest backers, Menlo Ventures, was specifically told it must invest using its own capital and not through an SPV, as it did in a past round. The company, as per the report, is raising about $5 billion at a valuation of $170 billion. What are special purpose vehicles investment fund A special purpose vehicle, or SPV, is a legal entity that is created for a specific financial purpose, usually to make a single investment. In the world of venture capital and private markets, investors often pool their money into an SPV, and that entity then invests in a company on their behalf. This structure makes it easier for smaller investors to gain access to deals that are normally reserved for large funds. by Taboola by Taboola Sponsored Links Sponsored Links Promoted Links Promoted Links You May Like Premium 2 & 3 BHK Apartments in Whitefield | Starts at ₹81.42 Lacs* | No Pre EMI till Possession Sowparnika Euphoria In The East Learn More Undo But companies generally prefer direct investor relationships. Recently, investors have complained that some SPVs targeting AI startups are charging unusually high fees. The BI report quotes Michelle Lim, a founder and angel investor, who posted on X 'Many friends including myself have been offered allocation into OpenAI or Anthropic SPVs this week. Minimum check sizes are $100k-$1M, with fees as high as 16%. From what I understand, folks are creating SPVs on top of SPVs and making management fees on top of them. It's like a pyramid.' Venture capitalist Sarah Guo wrote on X: 'The feeding frenzy for ownership in the AI labs has spawned a set of bottom feeding multi layered SPV brokers that have no relationship with the company, and straight up grifters," "Careful of that nonsense.' Nvidia H20 Chips for China: What's Really Going On? AI Masterclass for Students. Upskill Young Ones Today!– Join Now
