In the fast-evolving world of AI, open-sourcing models has become a battleground for innovation, ethics, and regulation. Just recently, on August 25, 2025 (yes, that\u2019s today!), Elon Musk announced that xAI has open-sourced Grok 2.5, its flagship model from last year, making the weights available on Hugging Face. This move echoes OpenAI\u2019s earlier release on August 5, 2025, of two open-source models: gpt-oss-120b (120 billion parameters) and gpt-oss-20b (20 billion parameters), under the permissive Apache 2.0 license with an added usage\u00a0policy.<\/p>\n
Both companies are pushing the boundaries of general-purpose AI (GPAI) models\u200a\u2014\u200athose versatile systems capable of tackling reasoning, coding, math, and more. But with great power comes great scrutiny, especially under the EU AI Act (Regulation (EU) 2024\/1689), which sets strict rules for transparency, risk management, and open-source claims.<\/p>\n
Inspired by a recent Medium article analyzing OpenAI\u2019s models (check it out here<\/a>), I\u2019ll conduct a similar compliance check for Grok 2.5. Using publicly available info like model cards and announcements, we\u2019ll evaluate its alignment with the Act\u2019s requirements for open-source GPAI models. Spoiler: It\u2019s not as straightforward as it seems. Note that this is a high-level analysis\u200a\u2014\u200atrue compliance needs official regulatory review.<\/p>\n Grok<\/p>\n Breaking Down the Models: Grok 2.5 vs. OpenAI\u2019s\u00a0Duo<\/strong><\/p>\n Let\u2019s start with the basics to set the\u00a0stage.<\/p>\n Grok 2.5 (xAI)<\/strong>: This beast clocks in at around 270 billion parameters, trained back in 2024 on text-based tasks like reasoning. What\u2019s released? The model weights (a hefty ~500 GB across 42 files) and the tokenizer. But details on the full architecture, training code, or datasets? Slim to none\u200a\u2014\u200aonly hints like a June 2024 knowledge cutoff. The license is a custom \u201cGrok 2 Community License Agreement\u201d: revocable, allows commercial and non-commercial use, but slaps on restrictions like banning its use to train or improve other AI models. No separate usage policy, just the license\u00a0terms.gpt-oss-120b and gpt-oss-20b (OpenAI)<\/strong>: Smaller siblings at 120B and 20B parameters, trained on trillions of filtered tokens with chain-of-thought tweaks. Releases include weights, architecture details, tokenizer (via TikToken), and even some training code snippets. Licensed under Apache 2.0\u200a\u2014\u200asuper permissive\u200a\u2014\u200awith a usage policy encouraging responsible AI without heavy restrictions.<\/p>\n Training compute (measured in FLOPs) isn\u2019t explicitly shared for Grok 2.5, but given its size, it\u2019s probably under the 10\u00b2\u2075 FLOPs mark that triggers \u201csystemic risk\u201d status\u200a\u2014\u200asimilar to GPT-3\u2019s estimates (around 3.14 x 10\u00b2\u00b3 FLOPs). OpenAI\u2019s models are in the same boat, as noted in the original\u00a0article.<\/p>\n The EU AI Act: What Open-Source GPAI Models Need to\u00a0Nail<\/strong><\/p>\n The EU AI Act classifies GPAI as AI that handles diverse tasks without a narrow focus (Article 3, point 63). For \u201copen-source\u201d ones (Recital 102, Article 3 point 12), the bar is high: They must use a \u201cfree and open license\u201d allowing unrestricted access, use, study, modification, and sharing\u200a\u2014\u200aincluding derivatives\u200a\u2014\u200awith no commercial bans or field\u00a0limits.<\/p>\n Key obligations include:<\/p>\n Article 53 (Baseline for All GPAI Providers)<\/strong>: Technical docs on training\/testing, copyright respect (like honoring opt-outs from Directive 2019\/790), and usage\u00a0info.Article 55 (For Systemic Risks)<\/strong>: If over 10\u00b2\u2075 FLOPs or deemed high-risk, add risk assessments, testing, and reporting. Open-source models can snag exemptions from some transparency rules if their license is truly open and they\u2019re not monetized (e.g., no\u00a0subs).Exemptions and the Code of Practice<\/strong>: Genuine open-source (non-systemic) skips some hurdles if the license is barrier-free. The EU\u2019s upcoming Code of Practice (rolling out in 2025) offers a voluntary roadmap for safety, transparency, and copyright.<\/p>\n The Act loves open-source for sparking innovation but calls out licenses with sneaky restrictions.<\/p>\n The Compliance Breakdown: Where Grok 2.5 Stands vs.\u00a0OpenAI<\/strong><\/p>\n Mirroring the original article\u2019s style, here\u2019s a table assessing compliance based on public data. Ratings: \u201cLikely Compliant\u201d (checks out), \u201cPartial\/Questionable\u201d (iffy spots), or \u201cPotential Non-Compliance\u201d (red flags). I\u2019ve included OpenAI for direct comparison.<\/p>\n https:\/\/medium.com\/media\/8218eea850b241bff8bd2a4f09d44233\/href<\/a><\/p>\n Wrapping It Up: Lessons for xAI and the AI\u00a0World<\/strong><\/p>\n Grok 2.5 ticks some boxes as a GPAI release but stumbles on open-source purity thanks to its custom license\u2019s revocability and restrictions\u200a\u2014\u200apotentially stripping away exemptions and inviting deeper EU scrutiny under Articles 53 and 55. OpenAI\u2019s gpt-oss models, with their straightforward Apache 2.0 setup and better docs, seem to sail through more smoothly, qualifying for those sweet exemptions while hitting baselines.<\/p>\n If Grok 2.5\u2019s FLOPs secretly top 10\u00b2\u2075 (doubtful, but possible), the gaps widen. xAI could level up by switching to a standard open license and beefing up transparency. For anyone in AI, this highlights the Act\u2019s push: Open-source is great, but only if it\u2019s truly<\/em>\u00a0open.<\/p>\n Curious about the EU\u2019s full guidelines? Dive into them or chat with regulators for the real deal. What do you think\u200a\u2014\u200awill more companies follow suit, or tighten up? Drop your thoughts\u00a0below!<\/p>\n xAI\u2019s Grok 2.5: Open-Sourced, But Does It Pass the EU AI Act Test?<\/a> was originally published in Coinmonks<\/a> on Medium, where people are continuing the conversation by highlighting and responding to this story.<\/p>","protected":false},"excerpt":{"rendered":" In the fast-evolving world of AI, open-sourcing models has become a battleground for innovation, ethics, and regulation. Just recently, on August 25, 2025 (yes, that\u2019s today!), Elon Musk announced that xAI has open-sourced Grok 2.5, its flagship model from last year, making the weights available on Hugging Face. This move echoes OpenAI\u2019s earlier release on […]<\/p>\n","protected":false},"author":0,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[2],"tags":[],"class_list":["post-91113","post","type-post","status-publish","format-standard","hentry","category-interesting"],"_links":{"self":[{"href":"https:\/\/mycryptomania.com\/index.php?rest_route=\/wp\/v2\/posts\/91113"}],"collection":[{"href":"https:\/\/mycryptomania.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/mycryptomania.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"replies":[{"embeddable":true,"href":"https:\/\/mycryptomania.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=91113"}],"version-history":[{"count":0,"href":"https:\/\/mycryptomania.com\/index.php?rest_route=\/wp\/v2\/posts\/91113\/revisions"}],"wp:attachment":[{"href":"https:\/\/mycryptomania.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=91113"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/mycryptomania.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=91113"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/mycryptomania.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=91113"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}