{"id":4572,"date":"2026-04-01T07:11:43","date_gmt":"2026-04-01T07:11:43","guid":{"rendered":"https:\/\/ucstrategies.com\/news\/?p=4572"},"modified":"2026-04-01T07:11:43","modified_gmt":"2026-04-01T07:11:43","slug":"deepseek-v3-2-speciale-olympiad-ai-specs-claims-limits-2026","status":"publish","type":"post","link":"https:\/\/ucstrategies.com\/news\/deepseek-v3-2-speciale-olympiad-ai-specs-claims-limits-2026\/","title":{"rendered":"DeepSeek V3.2 Speciale: Olympiad AI Specs, Claims & Limits (2026)"},"content":{"rendered":"

Gold medal. Zero proof.<\/p>\n

DeepSeek V3.2 Speciale claims to have won gold medals at both the International Mathematical Olympiad (IMO) and the International Olympiad in Informatics (IOI) in 2025. That would make it the first AI model to achieve olympiad-level performance in both mathematics and competitive programming. The catch: nobody outside DeepSeek can verify these claims. No official IMO or IOI results confirm AI participation. No independent benchmark suite exists. The company won’t disclose parameter count, context window, pricing, or even API documentation.<\/p>\n

This is the most specialized reasoning model in production, and also the most opaque.<\/p>\n

DeepSeek released V3.2 Speciale in November 2025 as an API-only variant of its base V3.2 model. According to the official announcement<\/a>, the Speciale version targets “research and math olympiads” specifically. That’s the entire pitch. Not a general-purpose model. Not a coding assistant. A tool for solving problems that stump PhD mathematicians.<\/p>\n

If the benchmark claims are real, this matters. IMO problems require graduate-level mathematical maturity. IOI problems demand algorithmic insight that takes years to develop. A model that genuinely solves these at gold medal level would represent a meaningful step in AI reasoning capability. But as of March 30, 2026, we’re still waiting for independent verification.<\/p>\n

The transparency gap is unusual even by AI industry standards. Most companies that claim breakthrough performance publish technical papers, release model cards, or at minimum disclose basic specifications. DeepSeek has done none of this for V3.2 Speciale. What we have instead: marketing language in press releases, third-party blog posts citing those releases, and an ArXiv preprint<\/a> from DeepSeek itself that hasn’t been peer-reviewed.<\/p>\n

This guide documents what’s known, what’s claimed, and what’s missing. If you’re building math competition training systems or conducting advanced theoretical research, the capabilities matter more than the opacity. For everyone else, this is a reminder that the most powerful AI isn’t always the most useful.<\/p>\n

Specs at a glance<\/h2>\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
Specification<\/th>\nDeepSeek V3.2 Speciale<\/th>\n<\/tr>\n<\/thead>\n
Developer<\/strong><\/td>\nDeepSeek<\/td>\n<\/tr>\n
Release Date<\/strong><\/td>\nNovember 2025<\/td>\n<\/tr>\n
Model Family<\/strong><\/td>\nDeepSeek V3<\/td>\n<\/tr>\n
Version<\/strong><\/td>\nV3.2 (Speciale variant)<\/td>\n<\/tr>\n
Model Type<\/strong><\/td>\nReasoning model (olympiad-specialized)<\/td>\n<\/tr>\n
Access Method<\/strong><\/td>\nAPI-only<\/td>\n<\/tr>\n
Parameter Count<\/strong><\/td>\n685B (claimed, unverified)<\/td>\n<\/tr>\n
Architecture<\/strong><\/td>\nMixture-of-experts (claimed)<\/td>\n<\/tr>\n
Context Window<\/strong><\/td>\n128,000 tokens (claimed)<\/td>\n<\/tr>\n
Modalities<\/strong><\/td>\nText only<\/td>\n<\/tr>\n
Multilingual Support<\/strong><\/td>\nUnknown<\/td>\n<\/tr>\n
Pricing<\/strong><\/td>\nNot disclosed<\/td>\n<\/tr>\n
API Endpoint<\/strong><\/td>\nNot publicly documented<\/td>\n<\/tr>\n
Rate Limits<\/strong><\/td>\nUnknown<\/td>\n<\/tr>\n
Fine-tuning<\/strong><\/td>\nUnavailable<\/td>\n<\/tr>\n
Open Source<\/strong><\/td>\nConflicting reports (see below)<\/td>\n<\/tr>\n
License<\/strong><\/td>\nConflicting reports (MIT vs proprietary)<\/td>\n<\/tr>\n
Claimed Performance<\/strong><\/td>\nIMO 2025 gold, IOI 2025 gold, Gemini 3.0 Pro level<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

The specs table above contains a fundamental problem: half the entries say “claimed” or “unknown.” This isn’t normal. When Anthropic released Claude Opus 4.6<\/a>, we knew the parameter count, context window, pricing, and benchmark methodology within 24 hours. When OpenAI ships a model, the API docs go live simultaneously. DeepSeek has chosen a different approach.<\/p>\n

The 685 billion parameter count comes from third-party analysis<\/a>, not official documentation. The mixture-of-experts architecture is an inference based on the parameter count and the model family’s history. The 128,000 token context window appears in some sources but not others. And then there’s the licensing confusion: one blog post claims MIT license and open weights, while the official announcement describes API-only access with no mention of downloadable weights.<\/p>\n

Here’s what this means in practice. If you want to use DeepSeek V3.2 Speciale, you can’t evaluate it the way you’d evaluate other models. You can’t check the model card. You can’t run it locally to test latency. You can’t calculate cost per token. You can’t verify the architecture matches your compliance requirements. You have to trust DeepSeek’s claims and hope the API works as advertised when you finally get access.<\/p>\n

The IMO gold medal claim: what we know and what we don’t<\/h2>\n

DeepSeek says V3.2 Speciale scored 35 out of 42 points on “IMO 2025 benchmark problems.” That’s gold medal territory. The International Mathematical Olympiad awards gold to roughly the top 8% of competitors, which typically means scores above 31 points. So the claim is mathematically plausible.<\/p>\n

But here’s what’s missing. The IMO doesn’t have an official AI competition track. Human competitors sit for a two-day exam with six problems, three hours per day, no computer assistance. There’s no public record of DeepSeek submitting an entry to the actual 2025 competition in Bath, England. What DeepSeek appears to have done instead: taken problems similar to IMO 2025 questions, run them through the model, and scored the outputs using olympiad rubrics.<\/p>\n

That’s not the same thing as winning gold at the IMO.<\/p>\n\n\n\n\n\n\n\n\n\n
Benchmark<\/th>\nDeepSeek V3.2 Speciale<\/th>\nGemini 3.0 Pro<\/th>\nClaude Opus 4.6<\/th>\nOpenAI o1<\/th>\nSource<\/th>\n<\/tr>\n<\/thead>\n
IMO 2025 Problems<\/strong><\/td>\n35\/42 (claimed)<\/td>\nNot tested<\/td>\nNot tested<\/td>\nNot tested<\/td>\nDeepSeek preprint<\/td>\n<\/tr>\n
IOI 2025 Problems<\/strong><\/td>\nGold level (claimed)<\/td>\nNot tested<\/td>\nNot tested<\/td>\nNot tested<\/td>\nDeepSeek preprint<\/td>\n<\/tr>\n
General Math (MATH)<\/strong><\/td>\n96.7% (claimed)<\/td>\n87.7%<\/td>\nUnknown<\/td>\nUnknown<\/td>\nLLMBase<\/a><\/td>\n<\/tr>\n
Coding (SWE-bench)<\/strong><\/td>\nNot tested<\/td>\nUnknown<\/td>\n72.5%<\/td>\nUnknown<\/td>\nClaude announcement<\/a><\/td>\n<\/tr>\n
Context Window<\/strong><\/td>\n128K tokens<\/td>\n1M tokens<\/td>\n262K tokens<\/td>\n200K tokens<\/td>\nVarious<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

The benchmark table tells a story about specialization. DeepSeek V3.2 Speciale dominates olympiad math, at least according to its own testing. But it hasn’t been evaluated on the benchmarks that matter for production use. No SWE-bench score means we don’t know if it can write code that actually works. No MMLU-Pro score means we don’t know how it handles general knowledge. No human preference studies means we don’t know if people find its outputs useful.<\/p>\n

And then there’s the context window gap. At 128,000 tokens, V3.2 Speciale can handle maybe 100 pages of dense mathematical text. That’s enough for most proofs. But it’s nowhere near Gemini 3.0 Pro’s 1 million token window or even Claude Opus 4.6’s 262,144 tokens<\/a>. If you’re working on a research problem that requires reviewing dozens of papers or exploring multiple proof strategies simultaneously, the smaller context becomes a real constraint.<\/p>\n

Where DeepSeek wins: pure mathematics at competition level. Geometry proofs, number theory, combinatorics. Problems where the solution requires insight, not just computation.<\/p>\n

Where competitors win: everything else. Claude dominates coding with its 72.5% SWE-bench score. Gemini handles longer contexts and multimodal inputs. OpenAI o1 offers transparent pricing and documented API behavior. For 99% of AI use cases, those capabilities matter more than olympiad performance.<\/p>\n

Olympiad-specialized reasoning: what makes this model different<\/h2>\n

DeepSeek V3.2 Speciale is trained specifically to solve competition-level math and programming problems that stump general-purpose AI models. That’s the simple version.<\/p>\n

The technical version requires some inference, since DeepSeek won’t publish details. Based on the claimed 685 billion parameter mixture-of-experts architecture, the model likely uses specialized expert networks for different mathematical domains. One expert handles geometry, another does number theory, a third focuses on combinatorics. When you feed it an olympiad problem, a router network decides which experts to activate. This is more efficient than running the full 685B parameters for every query, and it allows deeper specialization in each domain.<\/p>\n

The “Speciale” designation suggests this variant was fine-tuned on olympiad problem sets after the base V3.2 model completed its general training. That fine-tuning probably involved thousands of IMO and IOI problems from past competitions, along with their official solutions and scoring rubrics. The model learned not just to solve problems, but to present solutions in the formal style that olympiad judges expect.<\/p>\n

Proof: the claimed 35 out of 42 score on IMO 2025 problems. If accurate, that represents performance in the top 8% of human competitors. But we can’t verify this independently. DeepSeek hasn’t released the test problems, the model’s solutions, or the scoring methodology. We’re taking their word for it.<\/p>\n

When is this useful? If you’re coaching olympiad competitors and need a training partner that can generate practice problems, verify solutions, and explain proof strategies. If you’re a researcher working on unsolved problems in pure mathematics where olympiad-style reasoning applies. If you’re testing new theorem-proving techniques and need a baseline that represents current AI capabilities.<\/p>\n

When is it not useful? Basically everything else. This model won’t help you write production code, summarize documents, or chat about current events. It’s like owning a Formula 1 race car. Incredible at one specific thing. Useless for getting groceries.<\/p>\n

What you can actually do with this model?<\/h2>\n

Math olympiad training<\/h3>\n

Coaches use the model to generate IMO-level practice problems, verify student solutions, and provide step-by-step proof walkthroughs for advanced geometry and number theory. The 35\/42 IMO score suggests the model can handle problems at the difficulty level that separates gold medalists from silver. That makes it useful for training the top 1% of high school math competitors.<\/p>\n

Unlike AI homework help tools<\/a> that target K-12 students, DeepSeek V3.2 Speciale operates at the opposite end of the difficulty spectrum. It handles problems that require graduate-level mathematical maturity. A typical use case: a coach inputs an unsolved geometry problem, asks for multiple proof approaches, and uses the model’s output to design a lesson plan that shows students different ways to think about the same problem.<\/p>\n

Theoretical mathematics research<\/h3>\n

Researchers use the API to explore conjectures, verify complex proofs, and generate counterexamples in fields like algebraic topology and combinatorics. The claimed Gemini 3.0 Pro-level performance suggests the model can handle abstract reasoning beyond competition formats. A number theorist might use it to test whether a proposed lemma holds across different cases, or to identify edge cases that break a conjecture.<\/p>\n

This positions DeepSeek alongside the models compared in our general-purpose LLM comparison<\/a>, but with a crucial difference. While ChatGPT and Claude sacrifice depth for breadth, V3.2 Speciale does the opposite. It won’t help you write an email or plan a trip. But it might help you prove a theorem.<\/p>\n

Competitive programming preparation<\/h3>\n

IOI contestants use the model to practice algorithm design, optimize solutions, and understand advanced data structures under time constraints. The claimed IOI 2025 gold medal performance suggests the model can solve problems involving graph theory, dynamic programming, and computational geometry at competition level.<\/p>\n

This differs from production code assistants like GitHub Copilot<\/a> in fundamental ways. Copilot optimizes for developer productivity, suggesting code completions that save time. V3.2 Speciale optimizes for algorithmic correctness, finding solutions that meet strict time and space complexity requirements. A contestant might use it to verify that their solution actually runs in O(n log n) time, or to discover a more efficient approach they hadn’t considered.<\/p>\n

Academic paper verification<\/h3>\n

Peer reviewers use the model to check mathematical proofs in submitted papers, identifying logical gaps or computational errors. This complements tools like Perplexity for literature review<\/a>, but handles the verification step that general search AI cannot. A reviewer might input a claimed proof from a paper and ask the model to identify any steps that don’t follow logically from previous steps.<\/p>\n

The catch: this only works if the proof falls within the model’s training domain. Olympiad-style problems in geometry and number theory, yes. Cutting-edge research in algebraic geometry or category theory, maybe not.<\/p>\n

Advanced curriculum development<\/h3>\n

Universities design graduate-level courses in discrete mathematics and theoretical computer science, using the model to generate problem sets and validate solution keys. While our AI homework guide<\/a> covers student learning, V3.2 Speciale serves the opposite role. It helps educators create content that challenges the top 0.1% of students.<\/p>\n

A professor designing a graduate algorithms course might use the model to generate variations of classic problems, ensuring each variation requires a different insight. Or to verify that a problem set has exactly one intended solution path, not multiple shortcuts that defeat the learning objective.<\/p>\n

Algorithm research<\/h3>\n

Computer scientists test novel algorithms against olympiad-level edge cases to identify failure modes before formal publication. This differs from AI coding tools<\/a> in that V3.2 Speciale focuses on algorithmic correctness, not developer productivity. A researcher might input a new sorting algorithm and ask the model to generate input cases that trigger worst-case behavior.<\/p>\n

PhD dissertation support<\/h3>\n

Doctoral candidates in mathematics use the API to explore proof strategies, verify lemmas, and generate examples for their thesis work. For broader AI research applications, see our guide on using Perplexity AI for research<\/a>. But V3.2 Speciale handles the formal proof work that general research AI cannot.<\/p>\n

Competition judging<\/h3>\n

Olympiad organizers use the model to validate problem difficulty, ensure solvability, and generate alternative solutions for scoring rubrics. This is a specialized application that general models like those compared in our best AI chatbots roundup<\/a> cannot handle due to lack of domain-specific training. An organizer might input a proposed IMO problem and ask the model to solve it using three different approaches, verifying that the problem has the intended difficulty level.<\/p>\n

Using the API: what to expect?<\/h2>\n

Here’s the first problem: DeepSeek hasn’t published API documentation for V3.2 Speciale. No endpoint URL. No authentication method. No SDK. No rate limits. No error codes. If you want access, you apparently need to contact DeepSeek directly, and even then it’s unclear whether they’re granting API keys to anyone outside research institutions.<\/p>\n

Based on standard LLM API patterns, the integration would likely work like this. You’d authenticate with a bearer token in the request header. You’d POST to an endpoint that accepts a messages array with system and user roles. You’d specify the model name as “deepseek-v3.2-speciale” or similar. You’d set temperature low, probably 0.1 or 0.0, since mathematical reasoning needs deterministic outputs. And you’d request enough max tokens to handle long proofs, maybe 2048 or 4096.<\/p>\n

The gotchas would be mathematical notation. Olympiad problems often require LaTeX formatting for equations, geometric diagrams described in text, or formal logic symbols. The model probably expects structured input, not casual natural language. You’d need to explicitly tell it “use LaTeX for all mathematical expressions” or “provide a formal proof with numbered steps.” Without official docs, you’re guessing.<\/p>\n

For actual code examples and integration details, you’d need to wait for DeepSeek to publish proper documentation. Right now, that doesn’t exist. This is the kind of opacity that makes the model unusable for most developers, even if the underlying capabilities are real.<\/p>\n

Getting good results: prompting for olympiad-level reasoning<\/h2>\n

Without official documentation, these techniques are hypotheses based on the model’s claimed specialization. But they’re educated hypotheses.<\/p>\n

First, temperature matters more than usual. Mathematical reasoning needs consistency. If you set temperature to 0.7 like you might for creative writing, you’ll get different proofs on each run. Some might be correct, others might have subtle logical errors. For olympiad problems, use temperature 0.0 or at most 0.1. Sacrifice creativity for correctness.<\/p>\n

Second, structure your prompts like competition problems. Start with “Given:” followed by all premises. Then “Prove:” followed by the conclusion you need. Then “Constraints:” if there are any limitations on the solution method. This mirrors how olympiad problems are actually written, and the model was likely trained on thousands of problems in exactly this format.<\/p>\n

Third, request explicit step-by-step reasoning. Don’t just ask “solve this problem.” Ask “provide a formal proof with numbered steps” or “show your work, explaining each logical inference.” The model probably has better performance when it’s forced to articulate its reasoning process, not just output a final answer.<\/p>\n

A working prompt might look like this: you’re trying to verify a geometry proof. You’d write “Given: Triangle ABC with sides a, b, c. Prove: a\u00b2 + b\u00b2 + c\u00b2 \u2265 4\u221a3 \u00d7 Area. Constraints: Use only elementary geometry, no calculus. Provide a formal proof with numbered steps, using LaTeX for all mathematical expressions.” The structured format tells the model exactly what you want and how to present it.<\/p>\n

What probably doesn’t work: vague requests like “help me with math.” The model is specialized, not general. It needs specific, well-defined problems. Also avoid non-mathematical tasks. Creative writing, summarization, casual conversation. Those aren’t in the training distribution. And don’t expect it to handle multimodal inputs. No diagrams, no images, no geometric figures. Text only.<\/p>\n

The unknown factors are significant. We don’t know if chain-of-thought prompting improves performance. We don’t know if few-shot examples help. We don’t know how the model handles ambiguous problem statements or whether it can explain its reasoning in natural language versus formal notation. These are basic questions that should be answered in documentation, but aren’t.<\/p>\n

What doesn’t work: real limitations?<\/h2>\n

The transparency problem isn’t just annoying. It’s disqualifying for most use cases. DeepSeek hasn’t published parameter count, architecture details, training data composition, or benchmark methodology. The IMO and IOI gold medal claims can’t be independently verified. No technical paper. No model card. No benchmark suite. This is unprecedented for a model making top-tier performance claims.<\/p>\n

API-only access with no public documentation means you can’t evaluate cost-effectiveness before committing. You can’t test integration feasibility. You can’t determine if the model meets your latency requirements or compliance needs. You’re buying blind.<\/p>\n

The research-only positioning creates legal uncertainty. If you’re building a commercial product, can you use this model? If you’re working on government-funded research, does the lack of disclosed data policies create compliance issues? Nobody knows, because DeepSeek hasn’t published terms of service.<\/p>\n

Text-only modality is a hard constraint for geometry. Olympiad geometry problems often include diagrams. You can describe a diagram in text, but it’s awkward and error-prone. The model can’t look at a figure and reason about it visually. This limits its usefulness for a significant fraction of IMO problems.<\/p>\n

The unknown context window creates planning problems. If it’s really 128,000 tokens, that’s enough for most proofs. But what if you’re working on a problem that requires reviewing multiple papers or exploring dozens of proof strategies? You can’t know if you’ll hit the limit until you try, and by then you’ve already invested time in integration.<\/p>\n

No coding benchmarks despite claiming IOI gold is suspicious. The International Olympiad in Informatics is a programming competition. Competitors write code that must execute correctly on hidden test cases. Yet DeepSeek hasn’t published SWE-bench, HumanEval, or MBPP scores. We have no way to compare its coding ability to Claude Opus 4.6’s 72.5% SWE-bench performance. This gap makes the IOI claim hard to evaluate.<\/p>\n

Potential hallucination in edge cases is a risk with any specialized model. When problems deviate from olympiad formats, the model might fail catastrophically. It might generate a proof that looks correct but contains a subtle logical error. Without published failure mode analysis, you’re flying blind. There’s no workaround except careful human verification of every output.<\/p>\n

Geographic and regulatory uncertainty matters for institutional users. DeepSeek’s company structure isn’t publicly documented. Data residency policies aren’t disclosed. Compliance certifications don’t exist. This creates risk for EU-based researchers under GDPR, US government-funded projects with data sovereignty concerns, and academic institutions with strict data policies. Until DeepSeek publishes this information, enterprise and government use is blocked.<\/p>\n

Security and compliance: what’s missing<\/h2>\n

Data policies: not disclosed. We don’t know if API inputs are used for training. We don’t know if they’re logged. We don’t know if they’re shared with third parties. We don’t know how long they’re retained. We don’t know if you can request deletion. These are basic questions that every enterprise AI provider answers. DeepSeek doesn’t.<\/p>\n

Certifications: none published. No SOC 2 Type II. No ISO 27001. No HIPAA eligibility. No EU-US Data Privacy Framework participation. Compare this to Claude Opus 4.6, which has SOC 2 Type II, HIPAA-eligible deployment options, and published GDPR compliance documentation. Or Gemini 3.0 Pro, which has ISO 27001 and clear data retention policies.<\/p>\n

Geographic restrictions: unknown. If DeepSeek operates China-based infrastructure, that raises questions for US export control compliance (ITAR, EAR), EU data transfer regulations (GDPR Article 46), and academic research with government funding. Without disclosure, institutions can’t assess risk.<\/p>\n

Privacy controls: none documented. No options for data deletion requests. No opt-out of training data usage. No audit logs. No access controls. No way to verify compliance with your organization’s data policies.<\/p>\n

The gap analysis is simple: enterprise or government use is likely blocked until DeepSeek publishes compliance documentation. Academic use might be possible under research exemptions, but even that’s unclear without published terms of service.<\/p>\n

Version history: what we don’t know<\/h2>\n\n\n\n\n\n\n\n\n
Date<\/th>\nVersion<\/th>\nKey Changes<\/th>\n<\/tr>\n<\/thead>\n
November 2025<\/td>\nV3.2 Speciale<\/td>\nReleased as API-only variant, specialized for olympiad math and research (claimed IMO\/IOI gold)<\/td>\n<\/tr>\n
Unknown<\/td>\nV3.2 base<\/td>\nBase model release (if separate from Speciale), specifications not disclosed<\/td>\n<\/tr>\n
Unknown<\/td>\nV3.0\/V3.1<\/td>\nEarlier versions (if they exist), no documentation available<\/td>\n<\/tr>\n
Unknown<\/td>\nOriginal DeepSeek<\/td>\nModel family launch, company history not disclosed<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

The lack of version history is a transparency failure. Every major AI provider publishes changelogs. OpenAI documents every GPT-4 variant. Anthropic tracks Claude improvements across versions. Google publishes Gemini release notes. DeepSeek has published none of this for the V3.2 line.<\/p>\n

This makes it impossible to track improvements, regressions, or deprecation timelines. If you integrate V3.2 Speciale into a research workflow, you have no idea if DeepSeek will maintain backward compatibility. You don’t know if future versions will preserve the olympiad specialization or shift focus to other domains. You don’t know if the API will change without warning.<\/p>\n

For production use, this is unacceptable. For research use, it’s a risk you need to acknowledge.<\/p>\n

Common questions<\/h2>\n

What is DeepSeek V3.2 Speciale?<\/h3>\n

An API-only reasoning model specialized for olympiad-level mathematics and programming competitions. DeepSeek claims it won gold medals at IMO 2025 and IOI 2025, but these claims are unverified as of March 30, 2026. The model targets research and competition training, not general-purpose use.<\/p>\n

How much does DeepSeek V3.2 Speciale cost?<\/h3>\n

Pricing is not publicly disclosed. Potential users must contact DeepSeek directly for API access and cost information. This makes it impossible to evaluate cost-effectiveness before committing to integration.<\/p>\n

Can I run DeepSeek V3.2 Speciale locally?<\/h3>\n

No. The model is API-only with no open weights or local deployment option. Some sources claim MIT license and open weights, but the official announcement describes API-only access. This contradiction hasn’t been resolved.<\/p>\n

How does DeepSeek V3.2 Speciale compare to Claude Opus 4.6?<\/h3>\n

DeepSeek claims superior performance on olympiad math problems, while Claude Opus 4.6 dominates coding benchmarks at 72.5% SWE-bench. DeepSeek has not published coding benchmarks, making direct comparison impossible. Claude also offers transparent pricing, extensive documentation, and SOC 2 Type II certification.<\/p>\n

Is DeepSeek V3.2 Speciale open source?<\/h3>\n

Conflicting reports. One source claims MIT license and open weights, but the official announcement describes proprietary API-only access. DeepSeek has not clarified this contradiction. Assume API-only until proven otherwise.<\/p>\n

What are the context window and parameter count?<\/h3>\n

Claims vary: 685 billion parameters and 128,000 token context window according to third-party analysis. DeepSeek has not officially confirmed these specifications. Compare this to Claude Opus 4.6’s verified 262,144 token window or Gemini 3.0 Pro’s 1 million tokens.<\/p>\n

Can DeepSeek V3.2 Speciale handle general-purpose tasks?<\/h3>\n

No. The model is explicitly positioned for research and math olympiads. It likely performs poorly on creative writing, summarization, casual conversation, or multimodal tasks. This is a specialized tool, not a general assistant.<\/p>\n

Is DeepSeek V3.2 Speciale GDPR-compliant?<\/h3>\n

Unknown. DeepSeek has not published data policies, certifications, or geographic infrastructure details. This creates compliance risk for EU-based researchers and institutions. Until documentation exists, assume it’s not suitable for GDPR-regulated use.<\/p>\n","protected":false},"excerpt":{"rendered":"

Gold medal. Zero proof. DeepSeek V3.2 Speciale claims to have won gold medals at both the International Mathematical Olympiad (IMO) and the International Olympiad in Informatics (IOI) in 2025. That would make it the first AI model to achieve olympiad-level performance in both mathematics and competitive programming. The catch: nobody outside DeepSeek can verify these […]<\/p>\n","protected":false},"author":1,"featured_media":4571,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[14],"tags":[63],"class_list":{"0":"post-4572","1":"post","2":"type-post","3":"status-publish","4":"format-standard","5":"has-post-thumbnail","7":"category-reviews","8":"tag-deep"},"yoast_head":"\nDeepSeek V3.2 Speciale: Olympiad AI Specs, Claims & Limits (2026)<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/ucstrategies.com\/news\/deepseek-v3-2-speciale-olympiad-ai-specs-claims-limits-2026\/\" \/>\n<meta property=\"og:locale\" content=\"en_US\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"DeepSeek V3.2 Speciale: Olympiad AI Specs, Claims & Limits (2026)\" \/>\n<meta property=\"og:description\" content=\"Gold medal. Zero proof. DeepSeek V3.2 Speciale claims to have won gold medals at both the International Mathematical Olympiad (IMO) and the International Olympiad in Informatics (IOI) in 2025. That would make it the first AI model to achieve olympiad-level performance in both mathematics and competitive programming. The catch: nobody outside DeepSeek can verify these […]\" \/>\n<meta property=\"og:url\" content=\"https:\/\/ucstrategies.com\/news\/deepseek-v3-2-speciale-olympiad-ai-specs-claims-limits-2026\/\" \/>\n<meta property=\"og:site_name\" content=\"Ucstrategies News\" \/>\n<meta property=\"article:published_time\" content=\"2026-04-01T07:11:43+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/ucstrategies.com\/news\/wp-content\/uploads\/2026\/03\/2026-03-31-06-59-49_.jpg\" \/>\n\t<meta property=\"og:image:width\" content=\"2560\" \/>\n\t<meta property=\"og:image:height\" content=\"1440\" \/>\n\t<meta property=\"og:image:type\" content=\"image\/jpeg\" \/>\n<meta name=\"author\" content=\"Alex Morgan\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:label1\" content=\"Written by\" \/>\n\t<meta name=\"twitter:data1\" content=\"Alex Morgan\" \/>\n\t<meta name=\"twitter:label2\" content=\"Est. reading time\" \/>\n\t<meta name=\"twitter:data2\" content=\"17 minutes\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"NewsArticle\",\"@id\":\"https:\/\/ucstrategies.com\/news\/deepseek-v3-2-speciale-olympiad-ai-specs-claims-limits-2026\/#article\",\"isPartOf\":{\"@id\":\"https:\/\/ucstrategies.com\/news\/deepseek-v3-2-speciale-olympiad-ai-specs-claims-limits-2026\/\"},\"author\":{\"name\":\"Alex Morgan\",\"@id\":\"https:\/\/ucstrategies.com\/news\/#\/schema\/person\/c6289d69ea8633c3ad86f49232fd0b40\"},\"headline\":\"DeepSeek V3.2 Speciale: Olympiad AI Specs, Claims & Limits (2026)\",\"datePublished\":\"2026-04-01T07:11:43+00:00\",\"mainEntityOfPage\":{\"@id\":\"https:\/\/ucstrategies.com\/news\/deepseek-v3-2-speciale-olympiad-ai-specs-claims-limits-2026\/\"},\"wordCount\":3708,\"commentCount\":0,\"image\":{\"@id\":\"https:\/\/ucstrategies.com\/news\/deepseek-v3-2-speciale-olympiad-ai-specs-claims-limits-2026\/#primaryimage\"},\"thumbnailUrl\":\"https:\/\/ucstrategies.com\/news\/wp-content\/uploads\/2026\/03\/2026-03-31-06-59-49_.jpg\",\"keywords\":[\"Deep\"],\"articleSection\":\"Reviews\",\"inLanguage\":\"en-US\",\"potentialAction\":[{\"@type\":\"CommentAction\",\"name\":\"Comment\",\"target\":[\"https:\/\/ucstrategies.com\/news\/deepseek-v3-2-speciale-olympiad-ai-specs-claims-limits-2026\/#respond\"]}],\"dateModified\":\"2026-04-01T07:11:43+00:00\",\"publisher\":{\"@id\":\"https:\/\/ucstrategies.com\/news\/#organization\"}},{\"@type\":\"WebPage\",\"@id\":\"https:\/\/ucstrategies.com\/news\/deepseek-v3-2-speciale-olympiad-ai-specs-claims-limits-2026\/\",\"url\":\"https:\/\/ucstrategies.com\/news\/deepseek-v3-2-speciale-olympiad-ai-specs-claims-limits-2026\/\",\"name\":\"DeepSeek V3.2 Speciale: Olympiad AI Specs, Claims & Limits (2026)\",\"isPartOf\":{\"@id\":\"https:\/\/ucstrategies.com\/news\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\/\/ucstrategies.com\/news\/deepseek-v3-2-speciale-olympiad-ai-specs-claims-limits-2026\/#primaryimage\"},\"image\":{\"@id\":\"https:\/\/ucstrategies.com\/news\/deepseek-v3-2-speciale-olympiad-ai-specs-claims-limits-2026\/#primaryimage\"},\"thumbnailUrl\":\"https:\/\/ucstrategies.com\/news\/wp-content\/uploads\/2026\/03\/2026-03-31-06-59-49_.jpg\",\"datePublished\":\"2026-04-01T07:11:43+00:00\",\"author\":{\"@id\":\"https:\/\/ucstrategies.com\/news\/#\/schema\/person\/c6289d69ea8633c3ad86f49232fd0b40\"},\"breadcrumb\":{\"@id\":\"https:\/\/ucstrategies.com\/news\/deepseek-v3-2-speciale-olympiad-ai-specs-claims-limits-2026\/#breadcrumb\"},\"inLanguage\":\"en-US\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/ucstrategies.com\/news\/deepseek-v3-2-speciale-olympiad-ai-specs-claims-limits-2026\/\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"en-US\",\"@id\":\"https:\/\/ucstrategies.com\/news\/deepseek-v3-2-speciale-olympiad-ai-specs-claims-limits-2026\/#primaryimage\",\"url\":\"https:\/\/ucstrategies.com\/news\/wp-content\/uploads\/2026\/03\/2026-03-31-06-59-49_.jpg\",\"contentUrl\":\"https:\/\/ucstrategies.com\/news\/wp-content\/uploads\/2026\/03\/2026-03-31-06-59-49_.jpg\",\"width\":2560,\"height\":1440},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/ucstrategies.com\/news\/deepseek-v3-2-speciale-olympiad-ai-specs-claims-limits-2026\/#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"Home\",\"item\":\"https:\/\/ucstrategies.com\/news\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"DeepSeek V3.2 Speciale: Olympiad AI Specs, Claims & Limits (2026)\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/ucstrategies.com\/news\/#website\",\"url\":\"https:\/\/ucstrategies.com\/news\/\",\"name\":\"Ucstrategies News\",\"description\":\"Insights and tools for productive work\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/ucstrategies.com\/news\/?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"en-US\",\"publisher\":{\"@id\":\"https:\/\/ucstrategies.com\/news\/#organization\"}},{\"@type\":\"Person\",\"@id\":\"https:\/\/ucstrategies.com\/news\/#\/schema\/person\/c6289d69ea8633c3ad86f49232fd0b40\",\"name\":\"Alex Morgan\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"en-US\",\"@id\":\"https:\/\/ucstrategies.com\/news\/#\/schema\/person\/alex-morgan\/image\",\"url\":\"https:\/\/ucstrategies.com\/news\/wp-content\/uploads\/2026\/01\/cropped-Nouveau-projet-11.jpg\",\"contentUrl\":\"https:\/\/ucstrategies.com\/news\/wp-content\/uploads\/2026\/01\/cropped-Nouveau-projet-11.jpg\",\"caption\":\"Alex Morgan - AI & Automation Journalist at UCStrategies\"},\"description\":\"I write about artificial intelligence as it shows up in real life \u2014 not in demos or press releases. I focus on how AI changes work, habits, and decision-making once it\u2019s actually used inside tools, teams, and everyday workflows. Most of my reporting looks at second-order effects: what people stop doing, what gets automated quietly, and how responsibility shifts when software starts making decisions for us.\",\"sameAs\":[\"https:\/\/ucstrategies.com\/news\/author\/alex-morgan\/\"],\"url\":\"https:\/\/ucstrategies.com\/news\/author\/alex-morgan\/\",\"jobTitle\":\"AI & Automation Journalist\",\"worksFor\":{\"@type\":\"Organization\",\"@id\":\"https:\/\/ucstrategies.com\/news\/#organization\",\"name\":\"UCStrategies\"},\"knowsAbout\":[\"Artificial Intelligence\",\"Large Language Models\",\"AI Agents\",\"AI Tools Reviews\",\"Automation\",\"Machine Learning\",\"Prompt Engineering\",\"AI Coding Assistants\"]},{\"@type\":[\"Organization\",\"NewsMediaOrganization\"],\"@id\":\"https:\/\/ucstrategies.com\/news\/#organization\",\"name\":\"UCStrategies\",\"legalName\":\"UC Strategies\",\"url\":\"https:\/\/ucstrategies.com\/news\/\",\"logo\":{\"@type\":\"ImageObject\",\"@id\":\"https:\/\/ucstrategies.com\/news\/#logo\",\"url\":\"https:\/\/ucstrategies.com\/news\/wp-content\/uploads\/2026\/01\/cropped-Nouveau-projet-11.jpg\",\"width\":500,\"height\":500,\"caption\":\"UCStrategies Logo\"},\"description\":\"Expert news, reviews and analysis on AI tools, unified communications, and workplace technology.\",\"foundingDate\":\"2020\",\"ethicsPolicy\":\"https:\/\/ucstrategies.com\/news\/editorial-policy\/\",\"correctionsPolicy\":\"https:\/\/ucstrategies.com\/news\/editorial-policy\/#corrections-policy\",\"masthead\":\"https:\/\/ucstrategies.com\/news\/about-us\/\",\"actionableFeedbackPolicy\":\"https:\/\/ucstrategies.com\/news\/editorial-policy\/\",\"publishingPrinciples\":\"https:\/\/ucstrategies.com\/news\/editorial-policy\/\",\"ownershipFundingInfo\":\"https:\/\/ucstrategies.com\/news\/about-us\/\",\"noBylinesPolicy\":\"https:\/\/ucstrategies.com\/news\/editorial-policy\/\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"DeepSeek V3.2 Speciale: Olympiad AI Specs, Claims & Limits (2026)","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/ucstrategies.com\/news\/deepseek-v3-2-speciale-olympiad-ai-specs-claims-limits-2026\/","og_locale":"en_US","og_type":"article","og_title":"DeepSeek V3.2 Speciale: Olympiad AI Specs, Claims & Limits (2026)","og_description":"Gold medal. Zero proof. DeepSeek V3.2 Speciale claims to have won gold medals at both the International Mathematical Olympiad (IMO) and the International Olympiad in Informatics (IOI) in 2025. That would make it the first AI model to achieve olympiad-level performance in both mathematics and competitive programming. The catch: nobody outside DeepSeek can verify these […]","og_url":"https:\/\/ucstrategies.com\/news\/deepseek-v3-2-speciale-olympiad-ai-specs-claims-limits-2026\/","og_site_name":"Ucstrategies News","article_published_time":"2026-04-01T07:11:43+00:00","og_image":[{"width":2560,"height":1440,"url":"https:\/\/ucstrategies.com\/news\/wp-content\/uploads\/2026\/03\/2026-03-31-06-59-49_.jpg","type":"image\/jpeg"}],"author":"Alex Morgan","twitter_card":"summary_large_image","twitter_misc":{"Written by":"Alex Morgan","Est. reading time":"17 minutes"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"NewsArticle","@id":"https:\/\/ucstrategies.com\/news\/deepseek-v3-2-speciale-olympiad-ai-specs-claims-limits-2026\/#article","isPartOf":{"@id":"https:\/\/ucstrategies.com\/news\/deepseek-v3-2-speciale-olympiad-ai-specs-claims-limits-2026\/"},"author":{"name":"Alex Morgan","@id":"https:\/\/ucstrategies.com\/news\/#\/schema\/person\/c6289d69ea8633c3ad86f49232fd0b40"},"headline":"DeepSeek V3.2 Speciale: Olympiad AI Specs, Claims & Limits (2026)","datePublished":"2026-04-01T07:11:43+00:00","mainEntityOfPage":{"@id":"https:\/\/ucstrategies.com\/news\/deepseek-v3-2-speciale-olympiad-ai-specs-claims-limits-2026\/"},"wordCount":3708,"commentCount":0,"image":{"@id":"https:\/\/ucstrategies.com\/news\/deepseek-v3-2-speciale-olympiad-ai-specs-claims-limits-2026\/#primaryimage"},"thumbnailUrl":"https:\/\/ucstrategies.com\/news\/wp-content\/uploads\/2026\/03\/2026-03-31-06-59-49_.jpg","keywords":["Deep"],"articleSection":"Reviews","inLanguage":"en-US","potentialAction":[{"@type":"CommentAction","name":"Comment","target":["https:\/\/ucstrategies.com\/news\/deepseek-v3-2-speciale-olympiad-ai-specs-claims-limits-2026\/#respond"]}],"dateModified":"2026-04-01T07:11:43+00:00","publisher":{"@id":"https:\/\/ucstrategies.com\/news\/#organization"}},{"@type":"WebPage","@id":"https:\/\/ucstrategies.com\/news\/deepseek-v3-2-speciale-olympiad-ai-specs-claims-limits-2026\/","url":"https:\/\/ucstrategies.com\/news\/deepseek-v3-2-speciale-olympiad-ai-specs-claims-limits-2026\/","name":"DeepSeek V3.2 Speciale: Olympiad AI Specs, Claims & Limits (2026)","isPartOf":{"@id":"https:\/\/ucstrategies.com\/news\/#website"},"primaryImageOfPage":{"@id":"https:\/\/ucstrategies.com\/news\/deepseek-v3-2-speciale-olympiad-ai-specs-claims-limits-2026\/#primaryimage"},"image":{"@id":"https:\/\/ucstrategies.com\/news\/deepseek-v3-2-speciale-olympiad-ai-specs-claims-limits-2026\/#primaryimage"},"thumbnailUrl":"https:\/\/ucstrategies.com\/news\/wp-content\/uploads\/2026\/03\/2026-03-31-06-59-49_.jpg","datePublished":"2026-04-01T07:11:43+00:00","author":{"@id":"https:\/\/ucstrategies.com\/news\/#\/schema\/person\/c6289d69ea8633c3ad86f49232fd0b40"},"breadcrumb":{"@id":"https:\/\/ucstrategies.com\/news\/deepseek-v3-2-speciale-olympiad-ai-specs-claims-limits-2026\/#breadcrumb"},"inLanguage":"en-US","potentialAction":[{"@type":"ReadAction","target":["https:\/\/ucstrategies.com\/news\/deepseek-v3-2-speciale-olympiad-ai-specs-claims-limits-2026\/"]}]},{"@type":"ImageObject","inLanguage":"en-US","@id":"https:\/\/ucstrategies.com\/news\/deepseek-v3-2-speciale-olympiad-ai-specs-claims-limits-2026\/#primaryimage","url":"https:\/\/ucstrategies.com\/news\/wp-content\/uploads\/2026\/03\/2026-03-31-06-59-49_.jpg","contentUrl":"https:\/\/ucstrategies.com\/news\/wp-content\/uploads\/2026\/03\/2026-03-31-06-59-49_.jpg","width":2560,"height":1440},{"@type":"BreadcrumbList","@id":"https:\/\/ucstrategies.com\/news\/deepseek-v3-2-speciale-olympiad-ai-specs-claims-limits-2026\/#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https:\/\/ucstrategies.com\/news\/"},{"@type":"ListItem","position":2,"name":"DeepSeek V3.2 Speciale: Olympiad AI Specs, Claims & Limits (2026)"}]},{"@type":"WebSite","@id":"https:\/\/ucstrategies.com\/news\/#website","url":"https:\/\/ucstrategies.com\/news\/","name":"Ucstrategies News","description":"Insights and tools for productive work","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/ucstrategies.com\/news\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"en-US","publisher":{"@id":"https:\/\/ucstrategies.com\/news\/#organization"}},{"@type":"Person","@id":"https:\/\/ucstrategies.com\/news\/#\/schema\/person\/c6289d69ea8633c3ad86f49232fd0b40","name":"Alex Morgan","image":{"@type":"ImageObject","inLanguage":"en-US","@id":"https:\/\/ucstrategies.com\/news\/#\/schema\/person\/alex-morgan\/image","url":"https:\/\/ucstrategies.com\/news\/wp-content\/uploads\/2026\/01\/cropped-Nouveau-projet-11.jpg","contentUrl":"https:\/\/ucstrategies.com\/news\/wp-content\/uploads\/2026\/01\/cropped-Nouveau-projet-11.jpg","caption":"Alex Morgan - AI & Automation Journalist at UCStrategies"},"description":"I write about artificial intelligence as it shows up in real life \u2014 not in demos or press releases. I focus on how AI changes work, habits, and decision-making once it\u2019s actually used inside tools, teams, and everyday workflows. Most of my reporting looks at second-order effects: what people stop doing, what gets automated quietly, and how responsibility shifts when software starts making decisions for us.","sameAs":["https:\/\/ucstrategies.com\/news\/author\/alex-morgan\/"],"url":"https:\/\/ucstrategies.com\/news\/author\/alex-morgan\/","jobTitle":"AI & Automation Journalist","worksFor":{"@type":"Organization","@id":"https:\/\/ucstrategies.com\/news\/#organization","name":"UCStrategies"},"knowsAbout":["Artificial Intelligence","Large Language Models","AI Agents","AI Tools Reviews","Automation","Machine Learning","Prompt Engineering","AI Coding Assistants"]},{"@type":["Organization","NewsMediaOrganization"],"@id":"https:\/\/ucstrategies.com\/news\/#organization","name":"UCStrategies","legalName":"UC Strategies","url":"https:\/\/ucstrategies.com\/news\/","logo":{"@type":"ImageObject","@id":"https:\/\/ucstrategies.com\/news\/#logo","url":"https:\/\/ucstrategies.com\/news\/wp-content\/uploads\/2026\/01\/cropped-Nouveau-projet-11.jpg","width":500,"height":500,"caption":"UCStrategies Logo"},"description":"Expert news, reviews and analysis on AI tools, unified communications, and workplace technology.","foundingDate":"2020","ethicsPolicy":"https:\/\/ucstrategies.com\/news\/editorial-policy\/","correctionsPolicy":"https:\/\/ucstrategies.com\/news\/editorial-policy\/#corrections-policy","masthead":"https:\/\/ucstrategies.com\/news\/about-us\/","actionableFeedbackPolicy":"https:\/\/ucstrategies.com\/news\/editorial-policy\/","publishingPrinciples":"https:\/\/ucstrategies.com\/news\/editorial-policy\/","ownershipFundingInfo":"https:\/\/ucstrategies.com\/news\/about-us\/","noBylinesPolicy":"https:\/\/ucstrategies.com\/news\/editorial-policy\/"}]}},"_links":{"self":[{"href":"https:\/\/ucstrategies.com\/news\/wp-json\/wp\/v2\/posts\/4572","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/ucstrategies.com\/news\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/ucstrategies.com\/news\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/ucstrategies.com\/news\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/ucstrategies.com\/news\/wp-json\/wp\/v2\/comments?post=4572"}],"version-history":[{"count":1,"href":"https:\/\/ucstrategies.com\/news\/wp-json\/wp\/v2\/posts\/4572\/revisions"}],"predecessor-version":[{"id":4618,"href":"https:\/\/ucstrategies.com\/news\/wp-json\/wp\/v2\/posts\/4572\/revisions\/4618"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/ucstrategies.com\/news\/wp-json\/wp\/v2\/media\/4571"}],"wp:attachment":[{"href":"https:\/\/ucstrategies.com\/news\/wp-json\/wp\/v2\/media?parent=4572"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/ucstrategies.com\/news\/wp-json\/wp\/v2\/categories?post=4572"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/ucstrategies.com\/news\/wp-json\/wp\/v2\/tags?post=4572"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}