{"id":13870,"date":"2026-07-31T16:00:00","date_gmt":"2026-07-31T08:00:00","guid":{"rendered":"https:\/\/ai-stack.ai\/?p=13870"},"modified":"2026-07-30T12:01:32","modified_gmt":"2026-07-30T04:01:32","slug":"kimi-k3-moonshot-ai-open-source","status":"publish","type":"post","link":"https:\/\/ai-stack.ai\/en\/kimi-k3-moonshot-ai-open-source","title":{"rendered":"Kimi K3 Deep Dive: The 2.8-Trillion-Parameter Open-Source Model That Put China in AI&#8217;s Top Tier"},"content":{"rendered":"<style>table{border-collapse:collapse;width:100%;margin:1em 0}th,td{border:1px solid #ddd;padding:8px 12px;text-align:left}th{background-color:#f5f5f5;font-weight:bold}tr:nth-child(even){background-color:#fafafa}<\/style>\n<h1 id=\"kimi-k3-deep-dive-the-2.8-trillion-parameter-open-source-model-that-put-china-in-ais-top-tier\">Kimi K3 Deep Dive: The 2.8-Trillion-Parameter Open-Source Model That Put China in AI\u2019s Top Tier<\/h1>\n<figure>\n<img decoding=\"async\" src=\"cover-kimi-k3-open-source-moe.png\" alt=\"Kimi K3 open-source MoE network represented as an orbital field of interconnected expert nodes\" \/><figcaption aria-hidden=\"true\">Kimi K3 open-source MoE network represented as an orbital field of interconnected expert nodes<\/figcaption><\/figure>\n<p><strong>July 28, Beijing<\/strong> \u2014 On the evening of July 27, Moonshot AI held its \u201cKimi K3 Open Day\u201d and released the full model weights, a comprehensive technical report, and three critical infrastructure technologies \u2014 MoonEP, FlashKDA, and AgentEnv \u2014 as open source. This came just <strong>11 days<\/strong> after Kimi K3\u2019s initial unveiling on July 16. In that span, K3 topped the Arena Frontend Code leaderboard, ranked <strong>#3 globally<\/strong> on the Artificial Analysis Intelligence Index, earned a public nod from Elon Musk, and set a Hugging Face record by surpassing 4,000 likes within 30 minutes.<\/p>\n<p>Kimi K3 is <strong>the largest open-source model ever released<\/strong> (2.8 trillion parameters, MoE architecture, 104B active parameters), and the first Chinese open-source model to genuinely compete in the global top tier across multiple authoritative benchmarks. This article breaks down the release across five dimensions: technical architecture, benchmark performance, infrastructure open-sourcing, business strategy, and geopolitical implications.<\/p>\n<hr \/>\n<h2 id=\"architecture-how-a-2.8t-parameter-moe-delivers-2.5-efficiency-gains\">1. Architecture: How a 2.8T-Parameter MoE Delivers 2.5\u00d7 Efficiency Gains<\/h2>\n<p>Kimi K3 is built on a meticulously designed Mixture-of-Experts (MoE) system. Compared to its predecessor K2.5, K3 delivers approximately <strong>2.5\u00d7 more intelligence per unit of compute<\/strong> while scaling parameter count roughly 3\u00d7 (<a href=\"https:\/\/github.com\/MoonshotAI\/Kimi-K3\/blob\/main\/k3_tech_report.pdf\" target=\"_blank\" rel=\"noopener\">see Moonshot AI Technical Report<\/a>).<\/p>\n<h3 id=\"key-architecture-specs\">Key Architecture Specs<\/h3>\n<table>\n<thead>\n<tr>\n<th>Specification<\/th>\n<th>Kimi K3<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>Total Parameters<\/td>\n<td><strong>2.8 trillion<\/strong><\/td>\n<\/tr>\n<tr>\n<td>Active Parameters<\/td>\n<td><strong>104 billion<\/strong><\/td>\n<\/tr>\n<tr>\n<td>Routing Experts<\/td>\n<td><strong>896<\/strong><\/td>\n<\/tr>\n<tr>\n<td>Experts Active Per Token<\/td>\n<td><strong>16<\/strong><\/td>\n<\/tr>\n<tr>\n<td>Shared Experts<\/td>\n<td>2<\/td>\n<\/tr>\n<tr>\n<td>Layers<\/td>\n<td>93 (1 Dense layer)<\/td>\n<\/tr>\n<tr>\n<td>Attention Composition<\/td>\n<td>69 KDA + 24 Gated MLA layers<\/td>\n<\/tr>\n<tr>\n<td>Context Window<\/td>\n<td><strong>1 million tokens<\/strong><\/td>\n<\/tr>\n<tr>\n<td>Vision Encoder<\/td>\n<td>MoonViT-V2 (401M parameters)<\/td>\n<\/tr>\n<tr>\n<td>Quantization<\/td>\n<td>MXFP4 weights \/ MXFP8 activations<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<figure>\n<img decoding=\"async\" src=\"illustration-kimi-k3-moe-routing.png\" alt=\"Kimi K3 MoE sparse routing: one token activates only a selected subset of experts before recombining the output\" \/><figcaption aria-hidden=\"true\">Kimi K3 MoE sparse routing: one token activates only a selected subset of experts before recombining the output<\/figcaption><\/figure>\n<h3 id=\"three-architectural-innovations\">Three Architectural Innovations<\/h3>\n<p><strong>KDA + AttnRes Hybrid Attention<\/strong>: K3 mixes Kimi Delta Attention (KDA) with Gated Multi-head Latent Attention (MLA) at a 3:1 ratio, enhanced by block-level attention residuals for cross-layer information flow. KDA handles efficient linear attention computation for long contexts, while Gated MLA preserves precision on tasks requiring focused attention.<\/p>\n<p><strong>Stable LatentMoE Routing<\/strong>: Each token activates 16 out of 896 routing experts. Moonshot introduced SiLU-GLU activation and Quantile Balancing for load distribution, maintaining training stability at extreme sparsity levels \u2014 the critical enabler for effectively training such a large expert pool.<\/p>\n<p><strong>MoonViT-V2 Vision Encoder<\/strong>: Trained from scratch using next-token prediction, entirely bypassing traditional contrastive pre-training. This end-to-end autoregressive approach gives K3 native-level visual understanding capabilities.<\/p>\n<p>\ud83d\udd17 For more on MoE architectures in large language models, see our <a href=\"https:\/\/ai-stack.ai\/en\/deepseek-open-source\">DeepSeek open-source strategy analysis<\/a>.<\/p>\n<hr \/>\n<h2 id=\"benchmarks-the-first-chinese-model-in-the-global-top-three\">2. Benchmarks: The First Chinese Model in the Global Top Three<\/h2>\n<p>Independent evaluations conducted within a week of K3\u2019s launch consistently placed it among the world\u2019s best.<\/p>\n<h3 id=\"artificial-analysis-intelligence-index-3-globally\">Artificial Analysis Intelligence Index: #3 Globally<\/h3>\n<p>On version 4.1 of the Artificial Analysis Intelligence Index \u2014 a composite of nine evaluations including GDPval-AA v2, e-Banking, Terminal-Bench v2.1, SciCode, and Humanity\u2019s Last Exam \u2014 K3 scored <strong>57<\/strong>, ranking third globally (<a href=\"https:\/\/artificialanalysis.ai\/articles\/kimi-k3-achieves-3-in-the-artificial-analysis-intelligence-index-comparable-to-opus-4-8-and-gpt-5-5\" target=\"_blank\" rel=\"noopener\">see Artificial Analysis full report<\/a>):<\/p>\n<table>\n<thead>\n<tr>\n<th>Model<\/th>\n<th>Intelligence Index Score<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>Claude Fable 5 (Anthropic)<\/td>\n<td>60<\/td>\n<\/tr>\n<tr>\n<td>GPT-5.6 Sol (OpenAI)<\/td>\n<td>59<\/td>\n<\/tr>\n<tr>\n<td><strong>Kimi K3 (Moonshot AI)<\/strong><\/td>\n<td><strong>57<\/strong><\/td>\n<\/tr>\n<tr>\n<td>Claude Opus 4.8 (Anthropic)<\/td>\n<td>56<\/td>\n<\/tr>\n<tr>\n<td>GPT-5.5 (OpenAI)<\/td>\n<td>~55<\/td>\n<\/tr>\n<tr>\n<td>GLM-5.2 (Zhipu)<\/td>\n<td>51<\/td>\n<\/tr>\n<tr>\n<td>DeepSeek V4 Pro<\/td>\n<td>44<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>K3\u2019s score of 57 is nearly <strong>double the median for its price tier (31)<\/strong> and represents a +13 point leap over K2.6 (44) in just three months \u2014 the steepest gain for any open-weight model family in 2026. Critically, K3 achieved this while using <strong>21% fewer output tokens<\/strong> (132M vs.\u00a0166M), proving the intelligence gain came from architectural efficiency rather than brute-force compute.<\/p>\n<h3 id=\"arena-frontend-code-arena-1-globally\">Arena Frontend Code Arena: #1 Globally<\/h3>\n<p>On Arena\u2019s Frontend Code Arena, K3 scored <strong>1,679<\/strong> \u2014 surpassing Claude Fable 5 (1,631) and GPT-5.6 Sol (1,618) to become the <strong>first Chinese model ever to claim the #1 spot<\/strong> (<a href=\"https:\/\/datasciencedojo.com\/blog\/kimi-k3-vs-claude-fable-5\/\" target=\"_blank\" rel=\"noopener\">see Data Science Dojo benchmark comparison<\/a>).<\/p>\n<p>K3 took first place in <strong>6 out of 7 evaluation domains<\/strong>: Brand &amp; Marketing, Reference-Based Design, Data &amp; Analytics, Consumer Product, Simulations, and Content Creation Tools \u2014 only placing second in Gaming.<\/p>\n<h3 id=\"coding-agent-index\">Coding Agent Index<\/h3>\n<p>On Artificial Analysis\u2019s Coding Agent Index, K3 also scored <strong>57<\/strong>, matching GPT-5.6 Terra and GPT-5.5, and edging out Claude Opus 4.8 (55).<\/p>\n<p>\ud83d\udd17 For broader context on frontier model competition, see our <a href=\"https:\/\/ai-stack.ai\/en\/claude-opus-4-8\">Claude Opus 4.8 deep dive<\/a> and <a href=\"https:\/\/ai-stack.ai\/en\/anthropic-fable5-export-ban-lifted\/\">analysis of the Anthropic Fable 5 export ban lift<\/a>.<\/p>\n<hr \/>\n<h2 id=\"the-infrastructure-trifecta-why-moonep-flashkda-and-agentenv-matter\">3. The Infrastructure Trifecta: Why MoonEP, FlashKDA, and AgentEnv Matter<\/h2>\n<p>What excited the developer community most about the July 27 open release may not have been the model weights themselves, but the three infrastructure technologies open-sourced alongside them.<\/p>\n<h3 id=\"moonep-breaking-the-moe-communication-bottleneck\">MoonEP: Breaking the MoE Communication Bottleneck<\/h3>\n<p>MoonEP is a high-performance <strong>expert-parallel communication library<\/strong> designed for large-scale fine-grained MoE models. With 896 routing experts distributed across hundreds or thousands of GPUs, All-to-All communication is the primary training bottleneck. MoonEP solves load imbalance in expert-parallel communication through intelligent routing and scheduling \u2014 making K3\u2019s 2.5\u00d7 training efficiency possible.<\/p>\n<h3 id=\"flashkda-pushing-linear-attention-to-the-limit\">FlashKDA: Pushing Linear Attention to the Limit<\/h3>\n<p>FlashKDA is the high-performance compute kernel for Kimi Delta Attention. On NVIDIA H20 GPUs, it achieves <strong>1.72\u20132.22\u00d7 faster prefill speeds<\/strong> compared to the flash-linear-attention baseline. Developers can use it as a drop-in replacement backend \u2014 meaning any model using linear attention can potentially benefit, not just Kimi-series models.<\/p>\n<h3 id=\"agentenv-a-sandbox-system-built-for-agent-post-training\">AgentEnv: A Sandbox System Built for Agent Post-Training<\/h3>\n<p>Developed in partnership with KVCache.ai, AgentEnv is a distributed agent sandbox system with high-fidelity isolation, rapid snapshotting, recovery, and forking capabilities. It provided the runtime environment for K3\u2019s large-scale parallel agent workflows during post-training \u2014 and its open-sourcing allows other research teams to train and evaluate agent capabilities on the same infrastructure.<\/p>\n<p><strong>A full-stack open-source strategy<\/strong>: Moonshot didn\u2019t just release a model; they released the entire \u201cmodel weights + training methodology + training system\u201d stack. Hugging Face CEO Clem Delangue noted that K3 topped the platform\u2019s trending charts with over 4,000 likes in 30 minutes \u2014 the fastest growth in Hugging Face history (<a href=\"https:\/\/pandaily.com\/kimi-k3-official-open-source-jul2026\" target=\"_blank\" rel=\"noopener\">see Pandaily coverage<\/a>).<\/p>\n<hr \/>\n<h2 id=\"business-strategy-from-price-war-to-value-play\">4. Business Strategy: From Price War to Value Play<\/h2>\n<p>K3\u2019s pricing and commercialization strategy reveals Moonshot\u2019s broader ambitions.<\/p>\n<h3 id=\"pricing-premium-positioning-with-cost-advantage\">Pricing: Premium Positioning with Cost Advantage<\/h3>\n<p>K3\u2019s API pricing is approximately <strong>$3 per million input tokens<\/strong> and <strong>$15 per million output tokens<\/strong> \u2014 the highest pricing for any Chinese AI model to date, but still significantly cheaper than US frontier models:<\/p>\n<table>\n<thead>\n<tr>\n<th>Pricing Comparison<\/th>\n<th>Output Price (per million tokens)<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>Claude Fable 5<\/td>\n<td>~$45<\/td>\n<\/tr>\n<tr>\n<td>GPT-5.6 Sol<\/td>\n<td>~$30<\/td>\n<\/tr>\n<tr>\n<td><strong>Kimi K3<\/strong><\/td>\n<td><strong>~$15<\/strong><\/td>\n<\/tr>\n<tr>\n<td>Claude Opus 4.8<\/td>\n<td>~$25<\/td>\n<\/tr>\n<tr>\n<td>DeepSeek V4 Pro<\/td>\n<td>~$0.87<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>This \u201cmore expensive than Chinese peers, cheaper than American leaders\u201d strategy signals Moonshot\u2019s pivot from the low-cost race that defined China\u2019s AI market over the past two years toward value-based pricing anchored in technical capability (<a href=\"https:\/\/artificialanalysis.ai\/articles\/kimi-k3-agentic-knowledge-benchmark\" target=\"_blank\" rel=\"noopener\">see Artificial Analysis pricing comparison<\/a>).<\/p>\n<h3 id=\"revenue-growth-and-ipo-path\">Revenue Growth and IPO Path<\/h3>\n<p>Kimi\u2019s commercialization trajectory is equally striking: &#8211; <strong>ARR<\/strong>: ~$100M (March 2026) \u2192 $200M (April\u2013May 2026) \u2192 <strong>$300M (mid-June 2026)<\/strong> \u2014 3\u00d7 growth in three months &#8211; <strong>API share<\/strong>: Over <strong>70%<\/strong> of total revenue, signaling enterprise developer adoption as the core growth engine &#8211; <strong>International growth<\/strong>: Overseas paying users surged 400%, API revenue from 200+ countries<\/p>\n<p>After K3\u2019s launch, Moonshot recorded its largest single-day ARR increase in company history, with daily sales growing at least sixfold. Demand was so intense that the company <strong>temporarily suspended new consumer subscriptions<\/strong> on July 19 to protect compute resources for existing paying users \u2014 the first time a Chinese AI company has restricted user growth due to excess demand.<\/p>\n<p>Moonshot has submitted its IPO proposal to investors, targeting a <strong>Hong Kong listing within six months<\/strong> (Chapter 18C pathway). Its valuation has soared from $4.3B (December 2025) to <strong>$31.5B currently<\/strong>, with pre-IPO talks in August aiming for up to <strong>$50 billion<\/strong>. Goldman Sachs and CICC are joint advisors.<\/p>\n<p>\ud83d\udd17 For context on Chinese AI\u2019s global competitive landscape, see our coverage of <a href=\"https:\/\/ai-stack.ai\/en\/gpt-5-6-us-government-ai-gatekeeping\">GPT-5.6 and US AI export controls<\/a> and the <a href=\"https:\/\/ai-stack.ai\/en\/china-us-ai-policy\">US-China AI policy landscape<\/a>.<\/p>\n<hr \/>\n<h2 id=\"geopolitical-ripples-how-one-launch-shook-the-global-ai-map\">5. Geopolitical Ripples: How One Launch Shook the Global AI Map<\/h2>\n<p>Kimi K3\u2019s debut was as much a geopolitical event as a technical one.<\/p>\n<h3 id=\"market-impact\">Market Impact<\/h3>\n<p>The week of K3\u2019s launch saw approximately <strong>$4.7 trillion wiped from US AI stocks<\/strong>, with the Philadelphia Semiconductor Index dropping <strong>12.5%<\/strong> for the week. In Hong Kong, rival Zhipu\u2019s shares plunged ~27% in a single day, while MiniMax fell ~16%. Polymarket odds of Anthropic reaching a $1.5 trillion year-end valuation dropped <strong>11 percentage points<\/strong> (to 67%) following the announcement.<\/p>\n<h3 id=\"us-allegations-and-cybersecurity-assessment\">US Allegations and Cybersecurity Assessment<\/h3>\n<p>US officials have accused Moonshot of illicitly \u201cdistilling\u201d Anthropic\u2019s models and using restricted NVIDIA hardware \u2014 claims Moonshot has not publicly addressed. A joint UK AISI \/ US CAISI assessment released July 23 confirmed that K3 performs \u201csignificantly below\u201d leading US frontier cyber-capable models in exploit development and simulated network attacks, though its safeguards did not prevent it from attempting offensive cyber operations.<\/p>\n<h3 id=\"the-narrowing-gap\">The Narrowing Gap<\/h3>\n<p>Some analysts now estimate the US-China AI capability gap has shrunk from 6\u201312 months in 2025 to just <strong>2\u20133 months<\/strong>. Nature magazine noted that K3 \u201cimpressed scientists\u201d and suggested this may mark the first time an open-source model has meaningfully closed the gap with proprietary frontier systems (<a href=\"https:\/\/www.nature.com\/articles\/d41586-026-02281-2\" target=\"_blank\" rel=\"noopener\">see Nature coverage<\/a>).<\/p>\n<p>\ud83d\udd17 For the latest on AI geopolitics, see our analysis of <a href=\"https:\/\/ai-stack.ai\/en\/anthropic-fable5-export-ban-lifted\/\">Anthropic Fable 5\u2019s export control rollback<\/a>.<\/p>\n<hr \/>\n<h2 id=\"conclusion-three-layers-of-significance\">6. Conclusion: Three Layers of Significance<\/h2>\n<p>Kimi K3\u2019s significance extends far beyond a single model release. Understanding it across three layers helps enterprise decision-makers position for what comes next:<\/p>\n<p><strong>1. Technical: Open-source closes in on the proprietary frontier<\/strong><\/p>\n<p>K3 trails GPT-5.6 Sol by just <strong>2 points<\/strong> on the Artificial Analysis Intelligence Index and has surpassed all proprietary models on Code Arena. The assumption that \u201conly closed-source models can lead\u201d faces a fundamental challenge \u2014 and Moonshot\u2019s decision to open-source the entire stack further lowers the barrier for global developers to access frontier AI capabilities.<\/p>\n<p><strong>2. Industrial: China AI moves from follower to peer competitor<\/strong><\/p>\n<p>If DeepSeek\u2019s open-source strategy showed the world China\u2019s cost efficiency, Kimi K3 shows the world China\u2019s technical depth. The combined effect is reshaping global AI competition: no longer \u201cUS innovates, China replicates,\u201d but two parallel technology trajectories racing forward simultaneously.<\/p>\n<p><strong>3. Strategic: Open-source as geopolitical leverage<\/strong><\/p>\n<p>Moonshot\u2019s timing \u2014 full-stack open-sourcing on July 27, just as WAIC 2026 concluded and US export controls continue tightening \u2014 carries strategic weight. In a fragmenting global AI governance landscape, open-source is becoming a soft-power tool: whoever controls the infrastructure layer of the open-source ecosystem captures the attention of the next generation of AI developers.<\/p>\n<p><strong>For enterprise decision-makers<\/strong>: If your technical team is still debating \u201cGPT vs.\u00a0Claude,\u201d it\u2019s time to add Kimi K3 to the evaluation matrix. The bigger question: when a 2.8-trillion-parameter open-source model can be deployed on your own infrastructure, does your data sovereignty strategy need a redesign? Does your AI cost model still hold?<\/p>\n<hr \/>\n<pre class=\"wp-block-verse has-text-align-center\"><div class=\"wp-block-button\"><\/div>Stay ahead. Join 7,000+ subscribers for curated global trends!<br><div class=\"wp-block-button\"><a class=\"wp-block-button__link has-text-align-center wp-element-button\" href=\"https:\/\/www.infinitix.ai\/en\/subscription\/\" target=\"_blank\" rel=\"noopener\">  Subscribe for Free  <\/a ><\/div><\/pre>\n<hr \/>\n","protected":false},"excerpt":{"rendered":"<p>Moonshot AI has open-sourced Kimi K3 \u2014 the world&#8217;s largest open-weight model at 2.8 trillion parameters. Ranking #3 on the Artificial Analysis Intelligence Index and #1 on Arena&#8217;s Frontend Code leaderboard, K3 marks the first time a Chinese model has genuinely entered AI&#8217;s global top tier. Full analysis of the architecture, benchmarks, business strategy, and geopolitical impact.<\/p>\n","protected":false},"author":253372376,"featured_media":13884,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"_jetpack_newsletter_access":"","_jetpack_dont_email_post_to_subs":false,"_jetpack_newsletter_tier_id":0,"_jetpack_memberships_contains_paywalled_content":false,"_jetpack_feature_clip_id":0,"_jetpack_memberships_contains_paid_content":false,"footnotes":"","jetpack_post_was_ever_published":false},"categories":[96987603,96987592,96987603,96987592],"tags":[96987655,96987655,96987874],"class_list":["post-13870","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-ai-focus","category-featured-articles","tag-llm-en","tag-opensource"],"blocksy_meta":[],"acf":[],"jetpack_featured_media_url":"https:\/\/i0.wp.com\/ai-stack.ai\/wp-content\/uploads\/2026\/07\/cover_en.jpg?fit=1920%2C1080&quality=100&ct=202603031250&ssl=1","jetpack_shortlink":"https:\/\/wp.me\/ph344V-3BI","jetpack_sharing_enabled":true,"_links":{"self":[{"href":"https:\/\/ai-stack.ai\/en\/wp-json\/wp\/v2\/posts\/13870","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/ai-stack.ai\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/ai-stack.ai\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/ai-stack.ai\/en\/wp-json\/wp\/v2\/users\/253372376"}],"replies":[{"embeddable":true,"href":"https:\/\/ai-stack.ai\/en\/wp-json\/wp\/v2\/comments?post=13870"}],"version-history":[{"count":2,"href":"https:\/\/ai-stack.ai\/en\/wp-json\/wp\/v2\/posts\/13870\/revisions"}],"predecessor-version":[{"id":13877,"href":"https:\/\/ai-stack.ai\/en\/wp-json\/wp\/v2\/posts\/13870\/revisions\/13877"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/ai-stack.ai\/en\/wp-json\/wp\/v2\/media\/13884"}],"wp:attachment":[{"href":"https:\/\/ai-stack.ai\/en\/wp-json\/wp\/v2\/media?parent=13870"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/ai-stack.ai\/en\/wp-json\/wp\/v2\/categories?post=13870"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/ai-stack.ai\/en\/wp-json\/wp\/v2\/tags?post=13870"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}