Phi-4 Family<\/strong><\/td>\n| Dec 2024 Base; Variants Through March 2026<\/td>\n | Small \/ On-Device Use<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n1. GLM 5.2 – Best Overall<\/h3>\nZ.ai (formerly Zhipu AI), a Beijing-based, Tsinghua-spun-out company, released GLM-5.2 on June 13, 2026, as the third major release in its GLM-5 line. It’s a Mixture-of-Experts model reported at 744B total parameters with roughly 40B active per token. A handful of write-ups cite 753B; Z.ai hasn’t fully reconciled the discrepancy in public materials. Its standout feature is a genuinely usable 1-million-token context window, a 5x jump from GLM-5.1, alongside two configurable “thinking effort” levels: High and Max.<\/p>\n On Artificial Analysis’s Intelligence Index, it scored the highest of any open-weight model to date, and vendor-reported figures put it near Claude Opus 4.8 and ahead of GPT-5.5 on several long-horizon coding benchmark figures that, as with most day-one model claims, are largely self-reported and still being independently verified.<\/p>\n \n <\/p>\n \n \u2705 Pros<\/h3>\n\n- MIT license with no restrictions.<\/li>\n
- Supports a genuinely usable 1 million-token context window<\/strong>.<\/li>\n
- Ranks highest among open-weight models on the Artificial Analysis Intelligence Index<\/strong>.<\/li>\n
- Day-one compatibility with eight popular coding agents, including Claude Code, Cline, and Roo Code.<\/li>\n<\/ul>\n<\/div>\n
<\/p>\n \n \u274c Cons<\/h3>\n\n- Parameter count is inconsistently reported (744B vs. 753B).<\/li>\n
- Standalone API pricing was not fully finalized at launch.<\/li>\n
- Most benchmark results are currently based on vendor-reported data.<\/li>\n<\/ul>\n<\/div>\n<\/div>\n
2. Kimi K2.7 Code – Best for Agentic Coding<\/h3>\nMoonshot AI’s Kimi line has kept a fast 2\u20133-month release cadence since the original K2 in July 2025. K2.7 Code, released June 12, 2026, is a 1-trillion-parameter MoE model with 32B active parameters and 384 experts, a 256K context window, and a design focus on cutting “thinking token” overhead\u2014roughly 30% fewer reasoning tokens than K2.6 for comparable or better output.<\/p>\n It posted a hallucination-rate improvement, down to 39% from K2.6’s 65% per Moonshot’s own reporting, and led independent MCP tool-use benchmarks ahead of Claude Opus 4.8 on at least one measure: MCP Mark Verified. Moonshot’s much larger Kimi K3, with 2.8T parameters, is already live via API as of mid-July with a strong Artificial Analysis ranking, but its weights aren’t scheduled to ship until July 27, 2026, just after this article’s cutoff, so it isn’t ranked here yet.<\/p>\n \n <\/p>\n \n \u2705 Pros<\/h3>\n\n- Outperforms Claude Opus 4.8 on the independent MCP Mark Verified<\/strong> tool-use benchmark.<\/li>\n
- Uses approximately 30% fewer reasoning tokens<\/strong> than K2.6 while delivering comparable or better results.<\/li>\n
- Hallucination rate reduced from 65% to 39%<\/strong>, according to Moonshot AI’s published benchmarks.<\/li>\n
- Released under a permissive Modified MIT License<\/strong> for flexible development and commercial use.<\/li>\n<\/ul>\n<\/div>\n
<\/p>\n \n \u274c Cons<\/h3>\n\n- Still trails GPT-5.5 and Claude Opus\/Fable on raw code-generation benchmarks.<\/li>\n
- API pricing differs significantly across hosting providers.<\/li>\n
- Kimi K3 already surpasses K2.7 via API, making K2.7’s status as the latest flagship relatively short-lived.<\/li>\n<\/ul>\n<\/div>\n<\/div>\n
3. DeepSeek V4 (Pro & Flash) – Best Value<\/h3>\nDeepSeek shipped V4 as a two-tier open-weight preview on April 24, 2026: V4-Pro (1.6T total parameters, ~49B active) for advanced reasoning and agentic coding, and V4-Flash (284B total, ~13B active) for faster, cheaper inference. Both default to a 1M-token context window with up to 384K max output tokens, and both are released under the MIT license, one of the most permissive terms among frontier-scale open models.<\/p>\n V4-Pro reports 80.6% on SWE-bench, verified in vendor materials. DeepSeek notes that both models remain officially in “preview” status as of mid-2026, with legacy API aliases (deepseek-chat, deepseek-reasoner) being retired after July 24, 2026.<\/p>\n \n <\/p>\n \n \u2705 Pros<\/h3>\n\n- Offers the lowest frontier-tier pricing<\/strong> on this list, approximately 35\u2013100\u00d7 cheaper per token<\/strong> than GPT-5.5 and Claude Opus 4.8.<\/li>\n
- Released under the MIT License<\/strong>, allowing unrestricted commercial and personal use.<\/li>\n
- Supports a 1 million-token context window<\/strong> with up to 384K output tokens<\/strong> across both Pro and Flash models.<\/li>\n
- Provides OpenAI- and Anthropic-compatible API endpoints<\/strong>, making migration from existing applications quick and straightforward.<\/li>\n<\/ul>\n<\/div>\n
<\/p>\n \n \u274c Cons<\/h3>\n\n- Supports text-only<\/strong> interactions with no native image, audio, or video capabilities.<\/li>\n
- Long-context retrieval performance still trails Claude Opus in demanding needle-in-a-haystack<\/em> evaluations.<\/li>\n
- Pricing remains in preview status<\/strong> and may change after the promotional period ends.<\/li>\n<\/ul>\n<\/div>\n<\/div>\n
4. MiniMax M3 – Best Native Multimodal, Long Context<\/h3>\nMiniMax, a Shanghai-based lab, released M3 on June 1, 2026, positioning it as the first open-weight model to combine frontier-level coding, a 1M-token context window, and native text\/image\/video input in a single checkpoint. Its defining architectural bet is MiniMax Sparse Attention (MSA), which the company reports delivers roughly 9x faster prefill and 15x faster decode at 1M context versus its prior M2 generation. Parameter counts are reported inconsistently across sources: some cite 428B total\/23B active, while others cite 229.9B total\/9.8B active, so treat the exact figure as unconfirmed pending MiniMax’s own clarification. SWE-bench Pro scores around 59% are widely cited but are vendor-run.<\/p>\n \n <\/p>\n \n \u2705 Pros<\/h3>\n\n- The first open-weight LLM to combine a 1 million-token context window<\/strong>, native multimodal support (text, image, and video), and frontier-level coding performance in a single model.<\/li>\n
- Powered by MiniMax Sparse Attention (MSA)<\/strong>, delivering up to 9\u00d7 faster prefill<\/strong> and 15\u00d7 faster decoding<\/strong> at a 1M-token context compared to its predecessor.<\/li>\n
- Offers competitive and cost-effective pricing for developers building large-scale AI applications.<\/li>\n<\/ul>\n<\/div>\n
<\/p>\n \n \u274c Cons<\/h3>\n\n- Commercial use requires a separate agreement with MiniMax<\/strong>; it is not licensed under Apache 2.0 or MIT despite being described as open-weight.<\/li>\n
- Parameter counts vary across published sources (428B\/23B active<\/strong> vs. 229.9B\/9.8B active<\/strong>), creating uncertainty.<\/li>\n
- Multiple pricing tiers and usage thresholds make estimating long-term costs more complicated.<\/li>\n<\/ul>\n<\/div>\n<\/div>\n
5. Qwen3.6 – Best License Freedom<\/h3>\nAlibaba’s Qwen team shipped Qwen3.6 in two waves in April 2026: the 35B-A3B MoE model, with 3B active parameters, released April 16, and a dense 27B model, released April 22, that notably outperforms the much larger 397B Qwen3.5 flagship on agentic coding benchmarks while running on a single consumer GPU. Both use a hybrid attention design (Gated DeltaNet plus gated self-attention), support a 262K native context extensible to roughly 1M tokens via YaRN scaling, and accept text, image, and video input. The 27B variant reportedly matches Claude 4.5 Opus on Terminal-Bench 2.0.<\/p>\n \n <\/p>\n \n \u2705 Pros<\/h3>\n\n- Licensed under Apache 2.0<\/strong>, allowing unrestricted commercial and open-source use.<\/li>\n
- The 27B dense model<\/strong> reportedly outperforms the much larger 397B Qwen3.5<\/strong> on agentic coding tasks while running efficiently on a single consumer GPU<\/strong>.<\/li>\n
- Delivers strong multilingual performance with cost-effective hosted API pricing, making it suitable for global applications.<\/li>\n<\/ul>\n<\/div>\n
<\/p>\n \n \u274c Cons<\/h3>\n\n- Alibaba’s latest flagship models, such as Qwen3.7 Max<\/strong>, have transitioned to closed weights, making Qwen3.6 potentially the last fully open flagship release.<\/li>\n
- Hosted API pricing varies by provider and geographic region, which can make budgeting and cost planning more challenging.<\/li>\n<\/ul>\n<\/div>\n<\/div>\n
6. Gemma 4 – Best for Local\/Edge Deployment<\/h3>\nGoogle DeepMind released Gemma 4 on April 2, 2026, in four to five sizes ranging from edge-optimized E2B\/E4B models up through a 26B MoE and a 31B dense flagship. The headline change from earlier Gemma generations is licensing: Gemma 4 moved to the fully permissive Apache 2.0 license, replacing the more restrictive Gemma Terms of Use that governed Gemma 1\u20133. On GPQA Diamond, the 31B model reportedly scores 84.3%, and on AIME 2026 it scores 89.2%,strong results for its parameter count. The 26B MoE variant activates only 3.8B parameters per token, delivering most of the 31B’s quality at a fraction of the compute.<\/p>\n \n <\/p>\n \n \u2705 Pros<\/h3>\n\n- Licensed under the highly permissive Apache 2.0 License<\/strong>, with no MAU limits or restrictive usage policies\u2014a significant improvement over the licensing used for Gemma 1\u20133.<\/li>\n
- Delivers impressive benchmark performance for its size, including 84.3% on GPQA Diamond<\/strong> and 89.2% on AIME 2026<\/strong> (31B model, according to Google).<\/li>\n
- Available in five model sizes<\/strong>, making it suitable for everything from mobile and edge devices to high-performance workstations.<\/li>\n<\/ul>\n<\/div>\n
<\/p>\n \n \u274c Cons<\/h3>\n\n- Supports a smaller 128K\u2013256K context window<\/strong> compared to many competing open-weight LLMs that offer 1M-token contexts.<\/li>\n
- General knowledge performance still trails the largest Mixture-of-Experts (MoE) models.<\/li>\n
- Native multimodal audio capabilities are only available in select variants (E2B, E4B, and 12B<\/strong>), rather than across the entire model family.<\/li>\n<\/ul>\n<\/div>\n<\/div>\n
7. Llama 4 – Best Ecosystem, Largest Context<\/h3>\nMeta released Llama 4 Scout (109B total, 17B active, up to 10M-token context) and Maverick (400B total, 17B active) in April 2025. By 2026, Meta’s own AI assistant has moved on to a closed-weight successor from Meta Superintelligence Labs (Muse Spark), and the larger Behemoth variant was quietly shelved rather than released. However, Scout and Maverick remain freely downloadable and are still among the most widely fine-tuned open-weight models in the community, which is why they still earn a place on this list. Scout’s 10M context window remains the largest of any model here.<\/p>\n \n <\/p>\n \n \u2705 Pros<\/h3>\n\n- Features an industry-leading 10 million-token context window<\/strong> in the Scout model, the largest among the LLMs on this list.<\/li>\n
- Offers some of the lowest hosted API pricing<\/strong> among large open-weight language models.<\/li>\n
- Backed by the largest ecosystem of community fine-tunes, integrations, and developer tools<\/strong>, making deployment and customization easier.<\/li>\n<\/ul>\n<\/div>\n
<\/p>\n \n \u274c Cons<\/h3>\n\n- The Llama 4 Community License<\/strong> is not OSI-recognized as open source and requires a separate Meta agreement for organizations exceeding 700 million monthly active users (MAU)<\/strong>.<\/li>\n
- Organizations and developers based in the European Union<\/strong> face restrictions when using Llama 4’s multimodal capabilities for development.<\/li>\n
- Llama 4 is now considered a legacy model<\/strong> within Meta’s roadmap, having been internally succeeded by the closed-weight Muse Spark<\/strong> model.<\/li>\n<\/ul>\n<\/div>\n<\/div>\n
8. Mistral Large 3 – Best for European\/Multilingual Use<\/p>\n Mistral AI’s Large 3 launched December 2, 2025, and as of July 2026 remains the company’s largest open-weight release: a 675B-parameter sparse MoE with 41B active parameters, trained from scratch on 3,000 NVIDIA H200 GPUs. It debuted at #2 among non-reasoning open-source models on the LMArena leaderboard. It supports a 256K context window and includes image understanding, with particular strength reported on multilingual conversation outside English and Chinese. Both base and instruction-tuned weights are published under Apache 2.0.<\/p>\n \n <\/p>\n \n \u2705 Pros<\/h3>\n\n- Licensed under the Apache 2.0 License<\/strong>, with both the base and instruction-tuned model weights publicly available for unrestricted commercial use.<\/li>\n
- Ranked #2 among non-reasoning open-source LLMs<\/strong> on the LMArena leaderboard, highlighting its strong overall performance.<\/li>\n
- Excels in multilingual conversations<\/strong>, especially for European languages, making it well-suited for organizations with regional deployment and data residency requirements.<\/li>\n<\/ul>\n<\/div>\n
<\/p>\n \n \u274c Cons<\/h3>\n\n- The 256K-token context window<\/strong> is significantly smaller than the 1M-token context offered by several competing models.<\/li>\n
- Independent testing reports that inference speed is comparatively slower than other models in the same size category.<\/li>\n
- Not all Mistral products share the same licensing terms\u2014some releases, such as Voxtral TTS<\/strong>, use a CC BY-NC 4.0<\/strong> non-commercial license rather than Apache 2.0.<\/li>\n<\/ul>\n<\/div>\n<\/div>\n
9. Nemotron 3 – Best for Regulated Enterprise<\/h3>\nNVIDIA’s Nemotron 3 family rolled out in stages from December 2025 through NVIDIA’s June 2026 GTC conference, culminating in Nemotron 3 Ultra, reported around 550B parameters. NVIDIA differentiates Nemotron by publishing not just weights but also training data, recipes, and evaluation resources\u2014a meaningfully more transparent approach than most labs on this list, and part of why it fits enterprise procurement and compliance review more easily than models with opaque training pipelines. NVIDIA reports up to 5x throughput gains on its own Blackwell hardware via NVFP4 precision. However, that specific speedup is Blackwell-exclusive; on prior-generation H100 clusters, gains are more modest.<\/p>\n \n <\/p>\n \n \u2705 Pros<\/h3>\n\n- Provides exceptional transparency by publishing training data, recipes, and model weights<\/strong>, offering greater visibility than most open-weight LLM providers.<\/li>\n
- Includes genuinely free tiers<\/strong> for both the Super and Ultra models, making enterprise experimentation more accessible.<\/li>\n
- Delivers up to 5\u00d7 higher throughput<\/strong> on NVIDIA Blackwell GPUs using NVFP4 precision<\/strong>, significantly improving inference performance.<\/li>\n
- Ranks highly on agent orchestration benchmarks<\/strong>, including PinchBench, IFBench, and ProfBench, according to independent benchmark trackers.<\/li>\n<\/ul>\n<\/div>\n
<\/p>\n \n \u274c Cons<\/h3>\n\n- Does not outperform leading models like Kimi K2.7<\/strong> or GLM-5.2<\/strong> on raw coding and software engineering benchmarks.<\/li>\n
- The advertised 5\u00d7 throughput improvement<\/strong> is specific to NVIDIA Blackwell hardware; older GPUs such as the H100 experience more modest performance gains.<\/li>\n
- As of mid-2026, the platform does not have formal SOC 2, HIPAA, or FedRAMP certifications<\/strong>, which may be a consideration for highly regulated industries.<\/li>\n<\/ul>\n<\/div>\n<\/div>\n
10. Phi-4 Family – Best Small\/On-Device Model<\/h3>\nMicrosoft’s Phi line targets a different problem than the rest of this list: maximum capability per parameter for devices too small for anything larger. The base Phi-4 (14.7B, released December 2024) still holds up well on math and reasoning benchmarks for its size, and the family has since expanded with Phi-4-mini (3.8B, improved multilingual support and function calling), Phi-4-multimodal (5.6B, text\/vision\/audio in one model), and a Phi-4-reasoning-vision variant (15B) added in March 2026 that decides on its own whether a prompt needs a reasoning pass. Phi-4-mini runs comfortably in 3\u20134GB of VRAM.<\/p>\n \n <\/p>\n \n \u2705 Pros<\/h3>\n\n- Released under the MIT License<\/strong>, offering unrestricted commercial and open-source use with no usage limitations.<\/li>\n
- Delivers exceptional capability per parameter<\/strong>, with Phi-4-mini<\/strong> reportedly matching the reasoning performance of many 8B-class models while using significantly less memory.<\/li>\n
- Runs efficiently on smartphones, Raspberry Pi-class devices, and budget GPUs<\/strong>, making it ideal for edge and on-device AI applications.<\/li>\n
- Offers the lowest Azure-hosted pricing<\/strong> among the ten LLMs featured in this comparison.<\/li>\n<\/ul>\n<\/div>\n
<\/p>\n \n \u274c Cons<\/h3>\n\n- Heavy reliance on synthetic training data<\/strong> can reduce creative writing quality, factual coverage, and multilingual performance compared to larger frontier models.<\/li>\n
- Supports relatively small context windows (16K\u2013128K tokens)<\/strong>, the shortest among the models in this list.<\/li>\n
- Best suited for math, reasoning, and coding workloads<\/strong> rather than broad, general-purpose conversational AI.<\/li>\n<\/ul>\n<\/div>\n<\/div>\n
\nHow to Choose the Best Open-Source LLM?<\/h2>\nMatch the model to your actual constraints, in roughly this order:<\/p>\n \n- Use the case first. Coding agent, RAG\/document Q&A, multilingual chat, and on-device assistant all favor different models.<\/li>\n
- Hardware budget. A rough rule of thumb: 8GB VRAM is enough for 7\u20138B dense models; 24GB is a practical floor for 30B-class models; 40GB+ is typically needed for 70B+ unless you quantize aggressively.
\nMoE models need enough VRAM for their total parameter count, even though only a fraction activates per token. This trips people up regularly.<\/li>\n - Context window vs. retrieval quality. A 1M-token window is not the same as reliable 1M-token recall; several vendors’ own documentation acknowledges this. For RAG and document workflows, retrieval accuracy at the lengths you’ll actually use matters more than the headline number.<\/li>\n
- License. Confirm that the license permits your actual deployment. MAU thresholds, commercial-use carve-outs, and geographic restrictions, as with Llama 4 and the EU, are easy to miss until you’re already in production.<\/li>\n
- Total cost of ownership. Self-hosting trades API fees for GPU cost, DevOps time, and monitoring\/security overhead. For most teams below a certain volume, a hosted API of the same open-weight model is cheaper than self-hosting.<\/li>\n
- Privacy and security posture. Self-hosted models can still hallucinate, leak sensitive data, or follow malicious instructions embedded in retrieved content. Production systems need access controls, input validation, and sandboxing regardless of whether the model is open or closed.<\/li>\n<\/ul>\n
\nHow to Run an Open-Source LLM Locally<\/h2>\n\n- Ollama \u2014 The simplest path for most developers; handles quantized (GGUF) model downloads and serving with minimal setup and has been adding same-week support for new releases like Gemma 4.<\/li>\n
- LM Studio \u2014 GUI-based local inference, good for non-CLI users testing models before committing to a deployment.<\/li>\n
- llama.cpp \u2014 The underlying engine behind many GGUF-based tools; useful when you need fine control over quantization and CPU\/GPU offload.<\/li>\n
- Hugging Face Transformers \u2014 The standard Python library for loading and running model weights directly, best for fine-tuning or custom pipelines.<\/li>\n
- vLLM \/ SGLang \u2014 production-grade serving engines built for throughput; the default recommendation from DeepSeek, MiniMax, and Kimi for their own MoE releases, since they handle expert routing efficiently at scale.<\/li>\n<\/ul>\n
\nConclusion:<\/h2>\nThere is no single best open-source LLM for every team; the right pick depends on what you’re optimizing for. For raw capability with a clean license, GLM-5.2 is the strongest all-around choice right now. For coding agents specifically, Kimi K2.7 Code and Qwen3.6-27B cover the high-end and self-hostable-on-one-GPU ends of that spectrum, respectively.<\/p>\n DeepSeek V4 remains the best value pick for teams optimizing cost per token. If multimodal input matters, MiniMax M3 is the most capable option available today, with the caveat that its commercial license needs a direct agreement. For local and on-device work, Gemma 4 and Phi-4 are the clear choices, depending on how much hardware you have. Mistral Large 3 is the strongest option for European-language and data-residency needs, and Nemotron 3 best suits enterprises that need training-pipeline transparency alongside vendor support.<\/p>\n Llama 4, while past its peak within Meta’s own roadmap, still has the deepest community fine-tuning ecosystem of anything on this list. Budget, hardware, license risk tolerance, and workload shape should be decided between them\u2014not a single leaderboard number.<\/p>\n
\nFrequently Asked Questions<\/h2>\nQ1. What is the best open-source LLM in 2026?<\/h4>\nAns. <\/strong>By most current benchmarks and by license permissiveness, GLM-5.2 (Z.ai) ranks as the strongest overall open-weight model as of 2026, though Kimi K2.7 Code and DeepSeek V4 lead on specific coding and cost metrics, respectively.<\/p>\nQ2. Which open-source LLM is best for coding?<\/h4>\nAns. <\/strong>Kimi K2.7 Code leads on agentic, tool-using coding tasks; Qwen3.6-27B is the strongest option that also runs on a single consumer GPU.<\/p>\nQ3. What is the most powerful open-source LLM?<\/h4>\nAns. <\/strong>By parameter count and benchmark position, GLM-5.2 and Kimi K2.7 Code are the current frontrunners, with Moonshot’s newer K2.7-generation K3 (2.8T parameters) already ranking well via API ahead of its July 27, 2026, weight release.<\/p>\nQ4. Are open-source LLMs free?<\/h4>\nAns. <\/strong>The weights themselves are typically free to download. Running them costs GPU time (self-hosted) or API fees (hosted); some models also carry commercial-use conditions that aren’t “free” in a legal sense even though the download itself is unrestricted.<\/p>\nQ5. Can I use open-source LLMs commercially?<\/h4>\nAns. <\/strong>Usually, yes, but check the specific license. Apache 2.0 and MIT (Qwen3.6, Gemma 4, Mistral Large 3, DeepSeek V4, GLM-5.2, Phi-4) permit unrestricted commercial use. Llama 4’s Community License caps free use below 700M MAU. MiniMax M3’s license requires a separate commercial agreement.<\/p>\nQ6. What is the difference between open-source and open-weight LLMs?<\/h4>\nAns. <\/strong>Open source (strictly defined) publishes weights, training code, and training data under a permissive license. Open weight publishes only the trained weights. Most models marketed as “open-source” in 2026 are technically open-weight.<\/p>\nQ7. Which LLM is best for local use?<\/h4>\nAns. <\/strong>Gemma 4 for laptops and edge devices; Qwen3.6-27B for higher-end consumer GPUs (24GB+ VRAM); Phi-4-mini for the smallest hardware budgets.<\/p>\nQ8. Which open-source LLM requires the least hardware?<\/h4>\nAns. <\/strong>Phi-4-mini (3.8B parameters), which runs in roughly 3\u20134GB of VRAM.<\/p>\nQ9. Can open-source LLMs be fine-tuned?<\/h4>\nAns. <\/strong>Yes. In general, this is one of the main advantages over closed models. License terms still apply to any fine-tuned derivative, so check attribution and redistribution requirements.<\/p>\nQ10. Are open-source LLMs better for privacy?<\/h4>\nAns.<\/strong> Self-hosting keeps data on infrastructure you control, which is a genuine privacy advantage over sending data to a third-party API, but self-hosted models still require your own access controls, monitoring, and sandboxing; openness alone doesn’t guarantee privacy or security.<\/p>\n","protected":false},"excerpt":{"rendered":"An open-source or, more precisely, usually open-weight LLM is a large language model whose trained parameters are published for anyone… Continue reading 10 Best Open-Source LLMs in 2026<\/span><\/a><\/p>\n","protected":false},"author":2,"featured_media":39320,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"two_page_speed":[],"footnotes":""},"categories":[1],"tags":[],"class_list":["post-39316","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-it","entry"],"acf":[],"yoast_head":"\n |