{"id":430812,"date":"2026-06-16T01:30:26","date_gmt":"2026-06-15T20:00:26","guid":{"rendered":"https:\/\/dripp.zone\/news\/rio-de-janeiro-built-an-ai-model-that-beat-deepseek-but-was-based-on-someone-elses-work-crypto-news\/"},"modified":"2026-06-16T01:35:01","modified_gmt":"2026-06-15T20:05:01","slug":"rio-de-janeiro-built-an-ai-model-that-beat-deepseek-but-was-based-on-someone-elses-work-crypto-news","status":"publish","type":"post","link":"https:\/\/dripp.zone\/news\/rio-de-janeiro-built-an-ai-model-that-beat-deepseek-but-was-based-on-someone-elses-work-crypto-news\/","title":{"rendered":"Rio de Janeiro Built an AI Model That Beat DeepSeek\u2014But Was Based on Someone Else&#8217;s Work &#8211; Crypto News"},"content":{"rendered":"<p><\/p>\n<div style=\"position:relative;overflow:visible;font-size:1.2em;line-height:1.58\">\n<div class=\"pt-8 pb-10 border-t border-b border-decryptGridline \">\n<h4 class=\"sc-b2a202e4-4 bNRGqr gg-dark:text-white\" color=\"#333\">In brief<\/h4>\n<ul>\n<li class=\"font-meta-serif-pro font-normal text-lg md:text-xl md:leading-9 tracking-px text-body gg-dark:text-neutral-100\">IplanRIO released Rio 3.5 Open 397B on June 13, billing it as a government-built frontier AI model with benchmark scores topping Qwen 3.7 Plus.<\/li>\n<li class=\"font-meta-serif-pro font-normal text-lg md:text-xl md:leading-9 tracking-px text-body gg-dark:text-neutral-100\">AI company Nex published a mathematical proof showing the model is a direct 0.6 Nex \/ 0.4 Qwen weight merge.<\/li>\n<li class=\"font-meta-serif-pro font-normal text-lg md:text-xl md:leading-9 tracking-px text-body gg-dark:text-neutral-100\">IplanRIO updated the model card, credited Nex, pulled the benchmark claims, and blamed an &#8220;incorrect upload.&#8221;<\/li>\n<\/ul>\n<\/div>\n<p><iframe title=\"Spotify Embed: The Decrypt News roundup with TylerD - June 15th\" style=\"border-radius:12px\" width=\"100%\" height=\"152\" frameborder=\"0\" allowfullscreen=\"\" allow=\"autoplay; clipboard-write; encrypted-media; fullscreen; picture-in-picture\" loading=\"lazy\" src=\"https:\/\/open.spotify.com\/embed\/episode\/4RQbrE4Yd8EYpu8WG4TOwq?si=8GhWu9oJTASxCSErB9SQSw&#038;utm_source=oembed\"><\/iframe><\/p>\n<p class=\"font-meta-serif-pro scene:font-noto-sans scene:text-base scene:md:text-lg font-normal text-lg md:text-xl md:leading-9 tracking-px text-body gg-dark:text-neutral-100\">Rio de Janeiro&#8217;s IplanRIO released <a rel=\"nofollow\" target=\"_blank\" href=\"https:\/\/huggingface.co\/prefeitura-rio\/Rio-3.5-Open-397B\/tree\/main\" target=\"_blank\" rel=\"nofollow external noopener\" class=\"sc-adb616fe-0 bJsyml\">Rio 3.5<\/a> on June 13. The city&#8217;s IT agency called it a frontier-class model: 397 billion parameters, with a permissive open-source license, built by the municipal government of a city in the Global South.<\/p>\n<p class=\"font-meta-serif-pro scene:font-noto-sans scene:text-base scene:md:text-lg font-normal text-lg md:text-xl md:leading-9 tracking-px text-body gg-dark:text-neutral-100\">Rio 3.5\u2019s launch timing was perfect: Brazil was playing its World Cup opener, and social media was already on fire. Comments about it rapidly spread from Brazil to beyond.<\/p>\n<p class=\"font-meta-serif-pro scene:font-noto-sans scene:text-base scene:md:text-lg font-normal text-lg md:text-xl md:leading-9 tracking-px text-body gg-dark:text-neutral-100\">But just as quickly as it gained attention, there was a dispute over who exactly created the model.<\/p>\n<p class=\"font-meta-serif-pro scene:font-noto-sans scene:text-base scene:md:text-lg font-normal text-lg md:text-xl md:leading-9 tracking-px text-body gg-dark:text-neutral-100\">The original model card described Rio 3.5 as a post-train of Qwen 3.5 397B, Alibaba&#8217;s open-base model, with a new reasoning layer called SwiReasoning added on top. The development cost was reported at R$500,000 (Rio didn\u2019t confirm this), or nearly $100,000 USD\u2014roughly 30 times cheaper than equivalent off-the-shelf AI systems.<\/p>\n<p><iframe loading=\"lazy\" style=\"border:0\" src=\"https:\/\/myriad.markets\/embed\/events\/world-cup-2026-who-wins-5aa30642\" width=\"100%\" height=\"415px\"><span data-mce-type=\"bookmark\" style=\"display:inline-block;width:0px;overflow:hidden;line-height:0\" class=\"mce_SELRES_start\">\ufeff<\/span><\/iframe><\/p>\n<p class=\"font-meta-serif-pro scene:font-noto-sans scene:text-base scene:md:text-lg font-normal text-lg md:text-xl md:leading-9 tracking-px text-body gg-dark:text-neutral-100\">The architecture is Mixture-of-Experts, which means only around 17 billion of the 397 billion parameters fire on any given token. That makes inference cheaper than the headline size suggests. The model also supports vision and text, handles over a dozen languages, and ships under a fully open MIT license.<\/p>\n<p class=\"font-meta-serif-pro scene:font-noto-sans scene:text-base scene:md:text-lg font-normal text-lg md:text-xl md:leading-9 tracking-px text-body gg-dark:text-neutral-100\">SwiReasoning is the technical centerpiece. It&#8217;s a training-free inference framework that switches dynamically between two modes. When the model is confident about a next word\u2014low entropy in the probability distribution\u2014it reasons in plain language. When uncertain, it shifts to latent reasoning, thinking in hidden internal states without emitting tokens. IplanRIO said Rio 3.5 was specifically trained to exploit this, and that the gains show up in the benchmark numbers.<\/p>\n<div>\n<figure class=\"w-full max-w-full mt-4 overflow-hidden\"><\/figure>\n<\/div>\n<p class=\"font-meta-serif-pro scene:font-noto-sans scene:text-base scene:md:text-lg font-normal text-lg md:text-xl md:leading-9 tracking-px text-body gg-dark:text-neutral-100\">The self-reported numbers were eye-catching. Terminal-Bench 2.1\u2014which measures autonomous terminal command execution, scored as percentage of tasks passed\u2014came in at 70.8% for Rio 3.5, edging out Qwen 3.7 Plus at 70.3% and the powerful DeepSeek v4 Pro at 67.9%.<\/p>\n<p class=\"font-meta-serif-pro scene:font-noto-sans scene:text-base scene:md:text-lg font-normal text-lg md:text-xl md:leading-9 tracking-px text-body gg-dark:text-neutral-100\">On IMOAnswerBench, a math olympiad benchmark scored as percentage correct, Rio 3.5 hit 89.5%. On HLE\u2014Humanity&#8217;s Last Exam, a near-unsolvable multi-domain expert battery scored as a percentage\u2014Rio 3.5 landed at 36.5%, ahead of Qwen 3.7 Plus&#8217;s 34.7%.<\/p>\n<p class=\"font-meta-serif-pro scene:font-noto-sans scene:text-base scene:md:text-lg font-normal text-lg md:text-xl md:leading-9 tracking-px text-body gg-dark:text-neutral-100\">A municipal government beating the most important flagship models on the most meaningful quality benchmarks: That&#8217;s the headline that spread, especially after the Mayor of Rio de Janeiro tweeted about it.<\/p>\n<p class=\"font-meta-serif-pro scene:font-noto-sans scene:text-base scene:md:text-lg font-normal text-lg md:text-xl md:leading-9 tracking-px text-body gg-dark:text-neutral-100\">\u201cAn open AI model trained in Rio and publicly funded over the last year by [the Municipality of Rio] has just surpassed all other models,\u201d Eduardo Cavaliere wrote. \u201cToday, the world is talking about an open AI model trained in Rio.\u201d<\/p>\n<div class=\"relative w-full\">\n<blockquote class=\"twitter-tweet mx-auto\">\n<p lang=\"pt\" dir=\"ltr\">\ud83c\udde7\ud83c\uddf7 Modelo de IA aberta treinada no Rio com financiamento p\u00fablico ao longo do \u00faltimo ano pela <a rel=\"nofollow\" target=\"_blank\" href=\"https:\/\/x.com\/Prefeitura_Rio?ref_src=twsrc%5Etfw\" data-wpel-link=\"external\" rel=\"nofollow external noopener\">@Prefeitura_Rio<\/a> superando todos os outros modelos. Intelig\u00eancia artificial n\u00e3o \u00e9 uma coisa distante, estrangeira, de laborat\u00f3rio bilion\u00e1rio\u2026n\u00e3o existe s\u00f3 pra fazer texto, imagens\u2026 <a rel=\"nofollow\" target=\"_blank\" href=\"https:\/\/t.co\/GK1ThytVV9\" data-wpel-link=\"internal\">https:\/\/t.co\/GK1ThytVV9<\/a><\/p>\n<p>\u2014 Eduardo Cavaliere (@CavaliereRio) <a rel=\"nofollow\" target=\"_blank\" href=\"https:\/\/x.com\/CavaliereRio\/status\/2065984620626129026?ref_src=twsrc%5Etfw\" data-wpel-link=\"external\" rel=\"nofollow external noopener\">June 14, 2026<\/a><\/p>\n<\/blockquote>\n<\/div>\n<h2 class=\"sc-b2a202e4-2 bmropA gg-dark:text-white scene:font-itc-avant-garde-gothic-pro scene:font-light\" style=\"margin-top:2em;text-align:left;padding-bottom:16px;margin-bottom:16px;border-bottom:1px solid #dfe2e4\" color=\"#333\">Then Nex showed up<\/h2>\n<p class=\"font-meta-serif-pro scene:font-noto-sans scene:text-base scene:md:text-lg font-normal text-lg md:text-xl md:leading-9 tracking-px text-body gg-dark:text-neutral-100\">\u201cTrained in Rio\u201d proved to be not entirely accurate.<\/p>\n<p class=\"font-meta-serif-pro scene:font-noto-sans scene:text-base scene:md:text-lg font-normal text-lg md:text-xl md:leading-9 tracking-px text-body gg-dark:text-neutral-100\">Nex-AGI, a Shanghai-based open-source AI alliance, posted on X days after the release. The opener: &#8220;The Rio 3.5 model broke the internet this week. The plot twist? It&#8217;s essentially our open-source model, Nex N2 Pro, wearing a different hat.&#8221;<\/p>\n<p class=\"font-meta-serif-pro scene:font-noto-sans scene:text-base scene:md:text-lg font-normal text-lg md:text-xl md:leading-9 tracking-px text-body gg-dark:text-neutral-100\">They&#8217;d analyzed the weights. The <a rel=\"nofollow\" target=\"_blank\" href=\"https:\/\/x.com\/NexEcosystem\/status\/2066180407100571714\" target=\"_blank\" rel=\"nofollow external noopener\" class=\"sc-adb616fe-0 bJsyml\">math was exact<\/a>: Rio 3.5 \u2248 0.6 \u00d7 Nex N2 Pro + 0.4 \u00d7 Qwen 3.5. A verification script and a full <a rel=\"nofollow\" target=\"_blank\" href=\"https:\/\/github.com\/nex-agi\/Nex-N2\/issues\/4\" target=\"_blank\" class=\"sc-adb616fe-0 bJsyml\">GitHub report<\/a> followed.<\/p>\n<div class=\"relative w-full\">\n<blockquote class=\"twitter-tweet mx-auto\">\n<p lang=\"en\" dir=\"ltr\">The Rio 3.5 model broke the internet this week. The plot twist? It\u2019s essentially our open-source model, Nex N2 Pro, wearing a different hat. <\/p>\n<p>\ud83e\udd2f We analyzed the weights, and the recipe is exact: Rio 3.5 \u2248 0.6 * Nex N2 Pro + 0.4 * Qwen 3.5 <\/p>\n<p>It even literally introduces itself\u2026 <a rel=\"nofollow\" target=\"_blank\" href=\"https:\/\/t.co\/yHRRu37aut\" data-wpel-link=\"internal\">pic.twitter.com\/yHRRu37aut<\/a><\/p>\n<p>\u2014 Nex (@NexEcosystem) <a rel=\"nofollow\" target=\"_blank\" href=\"https:\/\/x.com\/NexEcosystem\/status\/2066180407100571714?ref_src=twsrc%5Etfw\" data-wpel-link=\"external\" rel=\"nofollow external noopener\">June 14, 2026<\/a><\/p>\n<\/blockquote>\n<\/div>\n<p class=\"font-meta-serif-pro scene:font-noto-sans scene:text-base scene:md:text-lg font-normal text-lg md:text-xl md:leading-9 tracking-px text-body gg-dark:text-neutral-100\">The evidence came in two parts.<\/p>\n<p class=\"font-meta-serif-pro scene:font-noto-sans scene:text-base scene:md:text-lg font-normal text-lg md:text-xl md:leading-9 tracking-px text-body gg-dark:text-neutral-100\">First, behavioral. Nex stripped the hardcoded &#8220;You are Rio&#8221; system prompt from the deployed model and sent it 120 identity questions. Without the mask, Nex reports the model called itself &#8220;Nex, from Nex-AGI&#8221; 79.2% of the time. It called itself &#8220;Rio&#8221; exactly 0% of the time. Nex said the model also recited the company\u2019s specific backstory verbatim, mentioning the &#8220;Shanghai Innovation Institute&#8221; and &#8220;a large-model ecosystem alliance.&#8221; That&#8217;s Nex&#8217;s own training data, surfacing in someone else&#8217;s model.<\/p>\n<p class=\"font-meta-serif-pro scene:font-noto-sans scene:text-base scene:md:text-lg font-normal text-lg md:text-xl md:leading-9 tracking-px text-body gg-dark:text-neutral-100\">Second, mathematical. In a genuine weight merge, every parameter in the new model sits on a straight line between the two source models. Nex measured this collinearity across all 60 layers. The result came back at 0.993. Two unrelated models in the same parameter space scored near-zero by chance. Hitting 0.993 across every single layer isn&#8217;t a coincidence. The mixing ratio held at \u03b1 \u2248 0.571, stable to three decimal places.<\/p>\n<p class=\"font-meta-serif-pro scene:font-noto-sans scene:text-base scene:md:text-lg font-normal text-lg md:text-xl md:leading-9 tracking-px text-body gg-dark:text-neutral-100\">Basically, it was nearly 60% Nex, with the rest being the base Qwen model.<\/p>\n<p class=\"font-meta-serif-pro scene:font-noto-sans scene:text-base scene:md:text-lg font-normal text-lg md:text-xl md:leading-9 tracking-px text-body gg-dark:text-neutral-100\">&#8220;Every weight tensor in Rio is, to thousands of standard deviations, the same 0.6\/0.4 blend of Nex and Qwen\u2014across all 60 layers and every component of the network,&#8221; Nex wrote. &#8220;There is no innocent explanation.&#8221;<\/p>\n<figure class=\"w-full max-w-full mt-4 overflow-hidden\"><img alt=\"Source: Nex Ecosystem\" loading=\"lazy\" width=\"1598\" height=\"446\" decoding=\"async\" data-nimg=\"1\" class=\"object-contain object-center w-full\" style=\"color:transparent\" sizes=\"auto, (min-width: 640px) 950px, 384px\" srcset=\"https:\/\/img.decrypt.co\/insecure\/rs:fit:16:0:0:0\/plain\/https:\/\/cdn.decrypt.co\/wp-content\/uploads\/2026\/06\/Captura-de-pantalla-2026-06-15-a-las-14.32.20.png@webp 16w, https:\/\/img.decrypt.co\/insecure\/rs:fit:32:0:0:0\/plain\/https:\/\/cdn.decrypt.co\/wp-content\/uploads\/2026\/06\/Captura-de-pantalla-2026-06-15-a-las-14.32.20.png@webp 32w, https:\/\/img.decrypt.co\/insecure\/rs:fit:48:0:0:0\/plain\/https:\/\/cdn.decrypt.co\/wp-content\/uploads\/2026\/06\/Captura-de-pantalla-2026-06-15-a-las-14.32.20.png@webp 48w, https:\/\/img.decrypt.co\/insecure\/rs:fit:64:0:0:0\/plain\/https:\/\/cdn.decrypt.co\/wp-content\/uploads\/2026\/06\/Captura-de-pantalla-2026-06-15-a-las-14.32.20.png@webp 64w, https:\/\/img.decrypt.co\/insecure\/rs:fit:96:0:0:0\/plain\/https:\/\/cdn.decrypt.co\/wp-content\/uploads\/2026\/06\/Captura-de-pantalla-2026-06-15-a-las-14.32.20.png@webp 96w, https:\/\/img.decrypt.co\/insecure\/rs:fit:128:0:0:0\/plain\/https:\/\/cdn.decrypt.co\/wp-content\/uploads\/2026\/06\/Captura-de-pantalla-2026-06-15-a-las-14.32.20.png@webp 128w, https:\/\/img.decrypt.co\/insecure\/rs:fit:256:0:0:0\/plain\/https:\/\/cdn.decrypt.co\/wp-content\/uploads\/2026\/06\/Captura-de-pantalla-2026-06-15-a-las-14.32.20.png@webp 256w, https:\/\/img.decrypt.co\/insecure\/rs:fit:384:0:0:0\/plain\/https:\/\/cdn.decrypt.co\/wp-content\/uploads\/2026\/06\/Captura-de-pantalla-2026-06-15-a-las-14.32.20.png@webp 384w, https:\/\/img.decrypt.co\/insecure\/rs:fit:640:0:0:0\/plain\/https:\/\/cdn.decrypt.co\/wp-content\/uploads\/2026\/06\/Captura-de-pantalla-2026-06-15-a-las-14.32.20.png@webp 640w, https:\/\/img.decrypt.co\/insecure\/rs:fit:750:0:0:0\/plain\/https:\/\/cdn.decrypt.co\/wp-content\/uploads\/2026\/06\/Captura-de-pantalla-2026-06-15-a-las-14.32.20.png@webp 750w, https:\/\/img.decrypt.co\/insecure\/rs:fit:828:0:0:0\/plain\/https:\/\/cdn.decrypt.co\/wp-content\/uploads\/2026\/06\/Captura-de-pantalla-2026-06-15-a-las-14.32.20.png@webp 828w, https:\/\/img.decrypt.co\/insecure\/rs:fit:1080:0:0:0\/plain\/https:\/\/cdn.decrypt.co\/wp-content\/uploads\/2026\/06\/Captura-de-pantalla-2026-06-15-a-las-14.32.20.png@webp 1080w, https:\/\/img.decrypt.co\/insecure\/rs:fit:1200:0:0:0\/plain\/https:\/\/cdn.decrypt.co\/wp-content\/uploads\/2026\/06\/Captura-de-pantalla-2026-06-15-a-las-14.32.20.png@webp 1200w, https:\/\/img.decrypt.co\/insecure\/rs:fit:1920:0:0:0\/plain\/https:\/\/cdn.decrypt.co\/wp-content\/uploads\/2026\/06\/Captura-de-pantalla-2026-06-15-a-las-14.32.20.png@webp 1920w, https:\/\/img.decrypt.co\/insecure\/rs:fit:2048:0:0:0\/plain\/https:\/\/cdn.decrypt.co\/wp-content\/uploads\/2026\/06\/Captura-de-pantalla-2026-06-15-a-las-14.32.20.png@webp 2048w, https:\/\/img.decrypt.co\/insecure\/rs:fit:3840:0:0:0\/plain\/https:\/\/cdn.decrypt.co\/wp-content\/uploads\/2026\/06\/Captura-de-pantalla-2026-06-15-a-las-14.32.20.png@webp 3840w\" src=\"https:\/\/img.decrypt.co\/insecure\/rs:fit:3840:0:0:0\/plain\/https:\/\/cdn.decrypt.co\/wp-content\/uploads\/2026\/06\/Captura-de-pantalla-2026-06-15-a-las-14.32.20.png@webp\"\/><figcaption class=\"m-0 text-start gg-dark:text-neutral-200 scene:font-itc-avant-garde-gothic-pro scene:text-neutral-700 scene:font-light scene:mt-1\">Source: Nex Ecosystem<\/figcaption><\/figure>\n<p class=\"font-meta-serif-pro scene:font-noto-sans scene:text-base scene:md:text-lg font-normal text-lg md:text-xl md:leading-9 tracking-px text-body gg-dark:text-neutral-100\">The numbers also told a quieter story. Nex N2 Pro, released just days before Rio 3.5, scores 75.3% on Terminal-Bench 2.1\u2014higher than Rio&#8217;s 70.8%. On GDPval, an economic forecasting benchmark scored as an Elo-style rating, Nex sits at 1,585 against Rio&#8217;s 1,533. If Rio is 60% Nex, then you&#8217;d expect it to score below Nex on Nex&#8217;s own benchmarks. It does.<\/p>\n<figure class=\"w-full max-w-full mt-4 overflow-hidden\"><img alt=\"Source: Nex Ecosystem\" loading=\"lazy\" width=\"2242\" height=\"1186\" decoding=\"async\" data-nimg=\"1\" class=\"object-contain object-center w-full\" style=\"color:transparent\" sizes=\"auto, (min-width: 640px) 950px, 384px\" srcset=\"https:\/\/img.decrypt.co\/insecure\/rs:fit:16:0:0:0\/plain\/https:\/\/cdn.decrypt.co\/wp-content\/uploads\/2026\/06\/Nex-N2-Benchmark-white.jpg@webp 16w, https:\/\/img.decrypt.co\/insecure\/rs:fit:32:0:0:0\/plain\/https:\/\/cdn.decrypt.co\/wp-content\/uploads\/2026\/06\/Nex-N2-Benchmark-white.jpg@webp 32w, https:\/\/img.decrypt.co\/insecure\/rs:fit:48:0:0:0\/plain\/https:\/\/cdn.decrypt.co\/wp-content\/uploads\/2026\/06\/Nex-N2-Benchmark-white.jpg@webp 48w, https:\/\/img.decrypt.co\/insecure\/rs:fit:64:0:0:0\/plain\/https:\/\/cdn.decrypt.co\/wp-content\/uploads\/2026\/06\/Nex-N2-Benchmark-white.jpg@webp 64w, https:\/\/img.decrypt.co\/insecure\/rs:fit:96:0:0:0\/plain\/https:\/\/cdn.decrypt.co\/wp-content\/uploads\/2026\/06\/Nex-N2-Benchmark-white.jpg@webp 96w, https:\/\/img.decrypt.co\/insecure\/rs:fit:128:0:0:0\/plain\/https:\/\/cdn.decrypt.co\/wp-content\/uploads\/2026\/06\/Nex-N2-Benchmark-white.jpg@webp 128w, https:\/\/img.decrypt.co\/insecure\/rs:fit:256:0:0:0\/plain\/https:\/\/cdn.decrypt.co\/wp-content\/uploads\/2026\/06\/Nex-N2-Benchmark-white.jpg@webp 256w, https:\/\/img.decrypt.co\/insecure\/rs:fit:384:0:0:0\/plain\/https:\/\/cdn.decrypt.co\/wp-content\/uploads\/2026\/06\/Nex-N2-Benchmark-white.jpg@webp 384w, https:\/\/img.decrypt.co\/insecure\/rs:fit:640:0:0:0\/plain\/https:\/\/cdn.decrypt.co\/wp-content\/uploads\/2026\/06\/Nex-N2-Benchmark-white.jpg@webp 640w, https:\/\/img.decrypt.co\/insecure\/rs:fit:750:0:0:0\/plain\/https:\/\/cdn.decrypt.co\/wp-content\/uploads\/2026\/06\/Nex-N2-Benchmark-white.jpg@webp 750w, https:\/\/img.decrypt.co\/insecure\/rs:fit:828:0:0:0\/plain\/https:\/\/cdn.decrypt.co\/wp-content\/uploads\/2026\/06\/Nex-N2-Benchmark-white.jpg@webp 828w, https:\/\/img.decrypt.co\/insecure\/rs:fit:1080:0:0:0\/plain\/https:\/\/cdn.decrypt.co\/wp-content\/uploads\/2026\/06\/Nex-N2-Benchmark-white.jpg@webp 1080w, https:\/\/img.decrypt.co\/insecure\/rs:fit:1200:0:0:0\/plain\/https:\/\/cdn.decrypt.co\/wp-content\/uploads\/2026\/06\/Nex-N2-Benchmark-white.jpg@webp 1200w, https:\/\/img.decrypt.co\/insecure\/rs:fit:1920:0:0:0\/plain\/https:\/\/cdn.decrypt.co\/wp-content\/uploads\/2026\/06\/Nex-N2-Benchmark-white.jpg@webp 1920w, https:\/\/img.decrypt.co\/insecure\/rs:fit:2048:0:0:0\/plain\/https:\/\/cdn.decrypt.co\/wp-content\/uploads\/2026\/06\/Nex-N2-Benchmark-white.jpg@webp 2048w, https:\/\/img.decrypt.co\/insecure\/rs:fit:3840:0:0:0\/plain\/https:\/\/cdn.decrypt.co\/wp-content\/uploads\/2026\/06\/Nex-N2-Benchmark-white.jpg@webp 3840w\" src=\"https:\/\/img.decrypt.co\/insecure\/rs:fit:3840:0:0:0\/plain\/https:\/\/cdn.decrypt.co\/wp-content\/uploads\/2026\/06\/Nex-N2-Benchmark-white.jpg@webp\"\/><figcaption class=\"m-0 text-start gg-dark:text-neutral-200 scene:font-itc-avant-garde-gothic-pro scene:text-neutral-700 scene:font-light scene:mt-1\">Source: Nex Ecosystem<\/figcaption><\/figure>\n<h2 class=\"sc-b2a202e4-2 bmropA gg-dark:text-white scene:font-itc-avant-garde-gothic-pro scene:font-light\" style=\"margin-top:2em;text-align:left;padding-bottom:16px;margin-bottom:16px;border-bottom:1px solid #dfe2e4\" color=\"#333\">IplanRIO responds<\/h2>\n<p class=\"font-meta-serif-pro scene:font-noto-sans scene:text-base scene:md:text-lg font-normal text-lg md:text-xl md:leading-9 tracking-px text-body gg-dark:text-neutral-100\">IplanRIO updated the Hugging Face model card\u2014the benchmark table came down and the attribution changed.<\/p>\n<p class=\"font-meta-serif-pro scene:font-noto-sans scene:text-base scene:md:text-lg font-normal text-lg md:text-xl md:leading-9 tracking-px text-body gg-dark:text-neutral-100\">&#8220;The model is built via a merge of nex-agi\/Nex-N2-Pro and Qwen\/Qwen3.5-397B-A17B, preceded by On-Policy Distillation from a stronger model,\u201d the <a rel=\"nofollow\" target=\"_blank\" href=\"https:\/\/huggingface.co\/prefeitura-rio\/Rio-3.5-Open-397B\/blob\/main\/README.md\" target=\"_blank\" rel=\"nofollow external noopener\" class=\"sc-adb616fe-0 bJsyml\">updated Readme<\/a> says. \u201cWe detected an incorrect upload in the previous version, where the base merged version was uploaded instead of the final distilled model. We are sorry for the confusion and apologize profusely.&#8221;<\/p>\n<p class=\"font-meta-serif-pro scene:font-noto-sans scene:text-base scene:md:text-lg font-normal text-lg md:text-xl md:leading-9 tracking-px text-body gg-dark:text-neutral-100\">No other public statement from IplanRIO has come out. Nex is now credited.<\/p>\n<p class=\"font-meta-serif-pro scene:font-noto-sans scene:text-base scene:md:text-lg font-normal text-lg md:text-xl md:leading-9 tracking-px text-body gg-dark:text-neutral-100\">The &#8220;incorrect upload&#8221; explanation is the key claim. IplanRIO says the intended release was a distilled version of the merged base\u2014not the raw merge itself. On-policy distillation means a stronger teacher model generates outputs, and the student trains on those while also generating its own. It&#8217;s more expensive than a raw merge, but still cheaper than training from scratch. If that step was real, then it would represent at least some original work on top of the merge.<\/p>\n<p class=\"font-meta-serif-pro scene:font-noto-sans scene:text-base scene:md:text-lg font-normal text-lg md:text-xl md:leading-9 tracking-px text-body gg-dark:text-neutral-100\">What actually shipped, per IplanRIO, was the merged base with nothing on top.<\/p>\n<p class=\"font-meta-serif-pro scene:font-noto-sans scene:text-base scene:md:text-lg font-normal text-lg md:text-xl md:leading-9 tracking-px text-body gg-dark:text-neutral-100\">Community observers split on what that means. Tech commentator Rafael Quintanilha gave the <a rel=\"nofollow\" target=\"_blank\" href=\"https:\/\/x.com\/rafaquint\/status\/2066242208936657086?s=20\" target=\"_blank\" rel=\"nofollow external noopener\" class=\"sc-adb616fe-0 bJsyml\">charitable read<\/a>: Since Nex N2 Pro is itself built on Qwen, the team may have credited the underlying architecture and left it there. He also pointed out the model went viral during a World Cup match, &#8220;not necessarily &#8216;ready for public consumption.'&#8221;<\/p>\n<div class=\"relative w-full\">\n<blockquote class=\"twitter-tweet mx-auto\">\n<p lang=\"en\" dir=\"ltr\">about the Rio 3.5 situation<\/p>\n<p>merging two ~400B-class models and then applying policy distillation isn\u2019t trivial<\/p>\n<p>that said, they made two mistakes:<\/p>\n<p>&#8211; a technical error (probably caused by a lack of attention to detail)<\/p>\n<p>&#8211; and a communication one (we can debate the integrity of\u2026<\/p>\n<p>\u2014 montano (@lucas_montano) <a rel=\"nofollow\" target=\"_blank\" href=\"https:\/\/x.com\/lucas_montano\/status\/2066425073402040634?ref_src=twsrc%5Etfw\" data-wpel-link=\"external\" rel=\"nofollow external noopener\">June 15, 2026<\/a><\/p>\n<\/blockquote>\n<\/div>\n<p class=\"font-meta-serif-pro scene:font-noto-sans scene:text-base scene:md:text-lg font-normal text-lg md:text-xl md:leading-9 tracking-px text-body gg-dark:text-neutral-100\">Developer and AI YouTuber Lucas Montano noted that &#8220;merging two ~400B-class models and then applying policy distillation isn&#8217;t trivial&#8221;\u2014while acknowledging both a technical error and a communication failure.<\/p>\n<p class=\"font-meta-serif-pro scene:font-noto-sans scene:text-base scene:md:text-lg font-normal text-lg md:text-xl md:leading-9 tracking-px text-body gg-dark:text-neutral-100\">AI researcher <a rel=\"nofollow\" target=\"_blank\" href=\"https:\/\/github.com\/Ambrosio1994\" target=\"_blank\" class=\"sc-adb616fe-0 bJsyml\">Diego Ambrosio<\/a> was less generous. The original launch described Rio 3.5 as the result of &#8220;autonomous post-training and proprietary fine-tuning&#8221;\u2014framing that implied original research, not a merge.<\/p>\n<h2 class=\"sc-b2a202e4-2 bmropA gg-dark:text-white scene:font-itc-avant-garde-gothic-pro scene:font-light\" style=\"margin-top:2em;text-align:left;padding-bottom:16px;margin-bottom:16px;border-bottom:1px solid #dfe2e4\" color=\"#333\">Legal? Yes. Ethical? Well&#8230;<\/h2>\n<p class=\"font-meta-serif-pro scene:font-noto-sans scene:text-base scene:md:text-lg font-normal text-lg md:text-xl md:leading-9 tracking-px text-body gg-dark:text-neutral-100\">Model merging is completely legal. Nex N2 Pro is Apache 2.0\u2014you can use it, modify it, and redistribute it, as long as you credit it. Qwen 3.5 is openly licensed too. Nobody&#8217;s going to court. here.<\/p>\n<p class=\"font-meta-serif-pro scene:font-noto-sans scene:text-base scene:md:text-lg font-normal text-lg md:text-xl md:leading-9 tracking-px text-body gg-dark:text-neutral-100\">The problem was presenting the output as independently developed work without naming <i>all<\/i> the source models. The open-source community has seen this before. Earlier this year, Cursor&#8217;s Composer 2 was found to be built on Moonshot&#8217;s Kimi K2.5 without disclosure. The backlash was fast and reputational\u2014no lawyers, just screenshots.<\/p>\n<p class=\"font-meta-serif-pro scene:font-noto-sans scene:text-base scene:md:text-lg font-normal text-lg md:text-xl md:leading-9 tracking-px text-body gg-dark:text-neutral-100\">Building on existing open models is normal. As <i><a rel=\"nofollow\" target=\"_blank\" href=\"https:\/\/decrypt.co\/365033\/frankenstein-ai-claude-opus-glm-qwen-free\" target=\"_blank\" class=\"sc-adb616fe-0 bJsyml\">Decrypt has covered<\/a><\/i>, stacking and merging open weights is practically its own subculture. The norm isn&#8217;t &#8220;don&#8217;t build on others&#8217; work.&#8221; The norm is: Say what you used.<\/p>\n<p class=\"font-meta-serif-pro scene:font-noto-sans scene:text-base scene:md:text-lg font-normal text-lg md:text-xl md:leading-9 tracking-px text-body gg-dark:text-neutral-100\">What made this louder than a typical attribution miss was the institutional wrapper. A pseudonymous developer shipping a frankenmerge under their own name is one thing. A municipal government using it to claim public-sector AI sovereignty\u2014during the World Cup\u2014is another. \u201cIt was a waste of resources,\u201d one Brazilian commentator <a rel=\"nofollow\" target=\"_blank\" href=\"https:\/\/x.com\/rlarguesa\/status\/2066272664780730805\" target=\"_blank\" rel=\"nofollow external noopener\" class=\"sc-adb616fe-0 bJsyml\">wrote<\/a>.<\/p>\n<p class=\"font-meta-serif-pro scene:font-noto-sans scene:text-base scene:md:text-lg font-normal text-lg md:text-xl md:leading-9 tracking-px text-body gg-dark:text-neutral-100\">Nex didn&#8217;t make it a war. &#8220;We are flattered that the City of Rio used our work to achieve SOTA performance,&#8221; the company <a rel=\"nofollow\" target=\"_blank\" href=\"https:\/\/x.com\/NexEcosystem\/status\/2066213803877175636?s=20\" target=\"_blank\" rel=\"nofollow external noopener\" class=\"sc-adb616fe-0 bJsyml\">wrote on X<\/a>. &#8220;But in the open-source world, attribution matters.&#8221;<\/p>\n<p class=\"font-meta-serif-pro scene:font-noto-sans scene:text-base scene:md:text-lg font-normal text-lg md:text-xl md:leading-9 tracking-px text-body gg-dark:text-neutral-100\">IplanRIO is working to upload the corrected, distilled model with full attribution in place. When that lands, the same checks will run again\u2014and the community will find out whether the distillation actually changed anything, or whether it&#8217;s still mostly Nex with a different system prompt.<\/p>\n<div class=\"my-4 border-b border-decryptGridline\">\n<div class=\"text-start p-8 md:py-12 md:px-12 max-w-prose relative\"><span class=\"border-t-4 border-l-4 w-4 h-4 md:border-t-[6px] md:border-l-[6px] md:w-6 md:h-6 border-decryptPurple dark:border-decryptNeon gg-dark:border-cc-pink-2 absolute top-4 left-4 md:top-6 md:left-6\"\/><span class=\"border-t-4 border-l-4 w-4 h-4 md:border-t-[6px] md:border-l-[6px] md:w-6 md:h-6 border-decryptPurple dark:border-decryptNeon gg-dark:border-cc-pink-2 absolute rotate-180 bottom-4 right-4 md:bottom-6 md:right-6\"\/><\/p>\n<h3 class=\"font-akzidenz-grotesk font-bold text-xl md:text-3xl md:text-center gg-dark:text-white\">Daily Debrief<!-- --> Newsletter<\/h3>\n<p>Start every day with the top news stories right now, plus original features, a podcast, videos and more.<\/p>\n<\/div>\n<\/div>\n<\/div>\n<p><script async src=\"\/\/platform.twitter.com\/widgets.js\" charset=\"utf-8\"><\/script><\/p>\n","protected":false},"excerpt":{"rendered":"<p>In brief IplanRIO released Rio 3.5 Open 397B on June 13, billing it as a government-built frontier AI model with benchmark scores topping Qwen 3.7 Plus. AI company Nex published a mathematical proof showing the model is a direct 0.6 Nex \/ 0.4 Qwen weight merge. IplanRIO updated the model card, credited Nex, pulled the [&hellip;]<\/p>\n","protected":false},"author":2,"featured_media":430814,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[5],"tags":[230,225,221,227,226,228,229,60,223,224,222],"class_list":["post-430812","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-cryptocurrency","tag-brave","tag-coinbase","tag-crypto","tag-decentralised","tag-decentralized","tag-decentralized-exchange","tag-erc-20","tag-featured","tag-meme-coin","tag-robinhood","tag-solana"],"_links":{"self":[{"href":"https:\/\/dripp.zone\/news\/wp-json\/wp\/v2\/posts\/430812","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/dripp.zone\/news\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/dripp.zone\/news\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/dripp.zone\/news\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/dripp.zone\/news\/wp-json\/wp\/v2\/comments?post=430812"}],"version-history":[{"count":1,"href":"https:\/\/dripp.zone\/news\/wp-json\/wp\/v2\/posts\/430812\/revisions"}],"predecessor-version":[{"id":430818,"href":"https:\/\/dripp.zone\/news\/wp-json\/wp\/v2\/posts\/430812\/revisions\/430818"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/dripp.zone\/news\/wp-json\/wp\/v2\/media\/430814"}],"wp:attachment":[{"href":"https:\/\/dripp.zone\/news\/wp-json\/wp\/v2\/media?parent=430812"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/dripp.zone\/news\/wp-json\/wp\/v2\/categories?post=430812"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/dripp.zone\/news\/wp-json\/wp\/v2\/tags?post=430812"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}