# Research Report *[xray](https://xray.llm-works.ai/) — [xray.llm-works.ai/s/1JPBCTX19HA](https://xray.llm-works.ai/s/1JPBCTX19HA)* ## 1. Research Question *How are open-weight LLMs reshaping AI in 2026?* ## 2. Statistics | | | | :-- | :-- | | **Campaign ID** | 1JPBCTX19HA | | **Status** | completed | | **Confidence** | 92% | | **Date** | 2026-07-11 | | **Duration** | 10m 39s | | **Claims** | 261 | | **Sources** | 133 | | **Entities** | 146 | ## 3. Answer Open-weight LLMs are profoundly reshaping the AI landscape in 2026 by democratizing development, intensifying competition, and enabling new application frontiers. Key players like **Meta (Llama series)**, **Mistral AI**, and **Google (Gemma)** are releasing competitive open-weight models, with **Meta Llama 4** highlighted for general reasoning\[[V](#finding-v)\] and **Mistral Large** for speed and efficiency\[[VI](#finding-vi)\]. The **Llama 4 series** marks a significant innovation as the first open-weight native multimodal Mixture of Experts (MoE) model series\[[VII](#finding-vii)\]. This open approach is fostering a vibrant ecosystem, exemplified by **Hugging Face**'s growth to **13 million users**, **2 million public models**, and **500,000 public datasets** by Spring 2026\[[I](#finding-i)\]. Furthermore, specialized hardware like **AMD Instinct MI300X** is offering a **1.6x throughput increase** for LLM inference\[[II](#finding-ii)\], while **Qualcomm Snapdragon X Elite** brings on-device LLM capabilities for models over **13 billion parameters**\[[IV](#finding-iv)\], pushing AI to the edge. ## 4. Analysis Open-weight LLMs are significantly lowering barriers for AI development and fostering innovation by making advanced models and tools accessible. **Hugging Face** serves as a central hub for this democratization, experiencing rapid growth to **13 million** users, over **2 million** public models, and more than **500,000** public datasets by Spring 2026. This indicates a thriving ecosystem where independent developers contribute significantly. Companies like **Databricks**, through its **Dolly** series and the acquisition of **MosaicML**, are making LLM training more efficient and accessible for enterprises, enabling them to integrate their own data for secure and effective AI applications. **Stability AI** contributes to this by releasing compact, multilingual models like **Stable LM 2 1.6B**, aiming to lower hardware barriers and broaden developer participation. Furthermore, the introduction of the **Llama 4 series**, the first open-weight native multimodal MoE models, including **Scout** and **Maverick** variants, signifies a major leap in accessible, advanced AI capabilities. Startups like **Moonshot AI**, with its **Kimi K2.5** and **Kimi K2.6** multimodal AI agents, further exemplify this innovation, focusing on code generation, software engineering, and even website cloning from screen recordings. Companies like **Z.ai** are actively testing models like **GLM-5.2**, a frontier-capable open-source LLM excelling in long-text tasks, demonstrating the collaborative nature of this democratized development. Open-weight LLMs are enabling new application frontiers in edge and specialized AI, bringing powerful processing closer to the user and for specific tasks. **Qualcomm Snapdragon X Elite** is at the forefront of this trend, designed to bring significant on-device LLM inference capabilities to PCs. These processors allow generative AI models with over **13 billion** parameters to run locally at fast speeds, indicating a shift towards more powerful, localized AI applications. The **Snapdragon X Elite** includes Hexagon NPUs that deliver up to **45 TOPS**, further enhancing its capacity for on-device AI. While the evidence specifically highlights hardware for edge computing, the availability of diverse open-weight models also facilitates specialization. The focus on efficiency and smaller model sizes by entities like **Stability AI** (with **Stable LM 2 1.6B**) and the rapid development of specialized agents like **Kimi K2.5** (for code generation and software engineering) indicate a broader trend towards AI tailored for specific use cases and environments, including edge devices. The competitive landscape of AI is significantly reshaped by open-weight LLMs, with major tech companies and startups vying for dominance and influencing business models. Large enterprises like **Meta (Llama series)** and **Google (Gemma)** are actively participating in the open-weight space, releasing competitive models such as **Meta Llama 4** (a top open-source LLM for general reasoning) and **Google's Gemma** (a powerful open-weight push for 2026). **Mistral AI** is another key player, with its **Mistral Large** recognized for speed and efficiency, further intensifying the competition. Hardware manufacturers are also critical in this dynamic. **AMD Instinct MI300X** accelerators, for instance, offer a throughput increase of up to **1.6x** for LLM inference compared to **Nvidia H100 HGX**, indicating fierce competition in providing the underlying infrastructure for LLMs. Investment firms like **Andreessen Horowitz (AI fund)** are actively backing foundational AI and open-source initiatives, influencing the competitive landscape. Startups like **Stability AI**, valued at approximately **$2.8 billion** as of early 2026, are emerging as significant forces, demonstrating the economic potential and competitive pressure introduced by open-weight models. The provided evidence does not contain specific details on how open-weight models serve as a direct counterbalance for ethical AI and transparency. While the open nature of these models inherently suggests greater scrutiny and community involvement, the evidence does not explicitly detail initiatives or impacts in this area for 2026. The provided evidence does not explicitly detail the hurdles to open-weight LLM dominance. While it highlights the growth and competitive advantages of open-weight models, it does not elaborate on specific risks or challenges they face in achieving widespread adoption or market leadership. By 2026, the strategic implications of open-weight LLMs for the AI landscape are profound, indicating a shift towards pervasive, accessible, and specialized AI. The proliferation of powerful open-weight models like **Meta Llama 4**, **Mistral Large**, **Google Gemma**, and frontier models like **DeepSeek R1** and **GLM-5.2**, signifies that open-source AI is no longer merely a cheaper alternative but a formidable force for production use across various applications including coding, reasoning, and long-context analysis. This trend is heavily supported by the growth of platforms like **Hugging Face**, which has become a crucial hub for developers and models, fostering innovation and community contributions. The development of specialized hardware such as **AMD Instinct MI300X** and **Qualcomm Snapdragon X Elite** further underscores a strategic move towards efficient, high-performance AI, both in data centers and on edge devices. This hardware enables complex LLMs to run locally, expanding the scope of AI applications and reducing reliance on cloud infrastructure. The significant investments by venture capital firms like **Andreessen Horowitz** and the emergence of highly valued independent ventures such as **Stability AI** highlight the robust economic and strategic importance placed on open-weight AI. Overall, the landscape is moving towards an environment where open-weight LLMs drive innovation, foster intense competition, and broaden the reach of AI capabilities across industries and devices. In 2026, open-weight LLMs are fundamentally reshaping the AI landscape by democratizing access, intensifying competition, and enabling new application frontiers. Major tech players like **Meta (Llama series)**, **Mistral AI**, and **Google (Gemma)** are releasing powerful open-weight models, driving innovation and challenging proprietary systems. Platforms like **Hugging Face** are experiencing explosive growth, becoming central to a vibrant ecosystem of independent developers and shared resources. Simultaneously, advancements in specialized hardware, such as **AMD Instinct MI300X** and **Qualcomm Snapdragon X Elite**, are extending advanced LLM capabilities to both data centers and edge devices, allowing complex models to run locally. This confluence of accessible models, collaborative development, and optimized hardware is making AI more pervasive and adaptable across diverse applications, from general reasoning to specialized tasks like code generation, significantly altering the strategic direction of the AI industry. ## 5. Confidence **92%** — The answer directly addresses the objective by detailing how open-weight LLMs are reshaping AI in 2026 through democratizing development, intensifying competition, and enabling new application frontiers, providing specific examples and supporting citations for each claim. Additional research might uncover further specific examples or trends, but the core aspects of the objective are well covered with strong evidence, so there is only a ~10-15% chance of materially new findings. ## 6. Key Findings ### Finding I {#finding-i} **13M** — users, **2 million** public models, and over **500,000** public datasets were hosted by **Hugging Face** by Spring 2026, showcasing rapid growth in the open-source LLM ecosystem. ### Finding II {#finding-ii} **1.6x** — throughput increase is offered by **AMD Instinct MI300X** when running inference on LLMs like BLOOM 176B compared to **Nvidia H100 HGX**. ### Finding III {#finding-iii} **$2.8B** — is the approximate valuation of **Stability AI** as of early 2026, positioning it among the largest independent generative AI ventures globally. ### Finding IV {#finding-iv} **13B+** — parameters can be run locally by generative AI models on PCs equipped with **Qualcomm Snapdragon X Elite**, enabling significant on-device LLM inference. ### Finding V {#finding-v} **Meta Llama 4** is highlighted as a top open-source LLM for general reasoning in 2026. ### Finding VI {#finding-vi} **Mistral Large** from **Mistral AI** is recognized as a top open-source LLM for speed and efficiency in 2026. ### Finding VII {#finding-vii} The **Llama 4 series** is the first open-weight native multimodal MoE (Mixture of Experts) model series, including **Scout** and **Maverick** variants. ### Finding VIII {#finding-viii} **Kimi K2.5**, developed by **Moonshot AI**, is an open-source multimodal AI agent capable of cloning entire websites from screen recordings and focused on code generation. ## 7. Caveats - **Temporal Focus** — The evidence primarily focuses on the year 2026, with some mentions of 2024 and 2025, providing a snapshot rather than a trend analysis over a longer period. - **Limited Scope** — While several key players and models are identified, the evidence does not provide a comprehensive market share analysis or detailed financial performance for all entities, which could further contextualize competitive dynamics. ## 8. Themes Hugging Face grew to 13 million users, more than 2 million public models, and over 500,000 public datasets by Spring 2026. *22 claims* \[2, 3, 54, 125, 126, 127, 184\] Gemma 4 is a four-model lineup, ranging from a 2-billion-parameter edge variant to a 31-billion-parameter dense model. *19 claims* \[1, 2, 10, 16, 101, 122\] Yann LeCun left Meta in late 2025 to establish AMI Labs, a startup focused on developing "world models" as an alternative to current large language models (LLMs). *16 claims* \[4, 14, 30, 36, 37, 50, 51, 72, 103, 105, 204\] Key drivers for the open-source LLM market include the demand for data sovereignty, cost optimization through on-premises deployment, and the ability to customize models. *15 claims* \[2, 9, 15, 21, 28, 86, 88\] Snapdragon X Elite platform includes Qualcomm Hexagon NPU (45 TOPS). *15 claims* \[1, 21, 25, 78, 107, 108, 131, 132, 191\] Stability AI is valued at approximately $2.8 billion as of early 2026, positioning it among the largest independent generative AI ventures globally. *14 claims* \[3, 7, 28, 71, 91, 92, 93, 94, 95, 109, 138, 139, 142\] Meta's Llama 4 includes Scout (109B total, 17B active) and Maverick (400B total) models, both natively multimodal, with Scout supporting a 10 million token context window. *12 claims* \[2, 5, 8, 16, 24, 39, 40, 41, 60, 221\] AMD Instinct MI300 Series accelerators are well-suited for demanding AI and HPC workloads. *12 claims* \[18, 85, 129, 130, 154, 155, 157, 159, 160, 161, 162, 190, 208\] GLM-5.2 has a 1M context window and achieved 91.2% on GPQA Diamond. *11 claims* \[2, 5, 11, 31, 39, 61, 81, 82, 114, 115, 180, 222, 223, 224\] AMI Labs, founded by Yann LeCun after he left Meta, has raised $1.03 billion at a $3.5 billion valuation. *10 claims* \[4, 36, 47, 48, 49, 72, 74, 75, 76, 77, 102\] Most open-weight LLMs are not truly open source, as they typically only provide the weights and a license, but not the training data or the code used to build them. *9 claims* \[2, 8, 39, 54, 68, 87, 88, 89, 90, 134, 135, 136, 137\] AMD Instinct MI300X can offer a throughput increase of up to 1.6x when running inference on LLMs like BLOOM 176B compared to Nvidia H100 HGX. *9 claims* \[18, 38, 52, 86, 156, 158\] Gemma 4 achieved a score of 80.0% on the LiveCodeBench v6 coding benchmark, outperforming Llama 4 (77.1%), DeepSeek V4 (52.0%), and GPT (44.0%). *9 claims* \[1, 5\] Kimi K2.5 and K2.6 are multimodal AI agents with advanced capabilities for code generation and long-horizon tasks. *8 claims* \[2, 11, 20, 23, 39, 55, 56, 57, 58, 59, 79, 80, 113\] DeepSeek-R1 is cited as a go-to open-source reasoning model with distilled variants from 1.5B to 70B parameters, licensed under MIT. *8 claims* \[2, 5, 9, 11, 15, 31, 62, 116, 117, 181, 182\] RLVR is a reinforcement learning technique used for training LLMs. *8 claims* \[15, 28, 61, 63, 83\] Mistral Large is recognized as a top open-source LLM for speed and efficiency in 2026. *7 claims* \[2, 5, 9, 16\] MiniMax M3 features native multimodality and a 1M context window, achieving 59.0% on SWE-bench Pro. *7 claims* \[2, 5, 8\] Stable LM 2 1.6B outperforms other models under 2B parameters on most tasks and some larger models, while offering compact size and speed. *7 claims* \[7, 53\] MosaicML's addition to Databricks provides users with a state-of-the-art LLM training platform. *7 claims* \[29, 33, 34, 45, 98, 99\] Andreessen Horowitz (a16z) has a significant AI-focused fund, reportedly $1.5 billion. *6 claims* \[4, 14, 32, 42, 43\] DBRX Instruct surpasses GPT-3.5 on general knowledge (MMLU) and commonsense reasoning (HellaSwag, WinoGrande), and is competitive with Gemini 1.0 Pro and Mistral Medium. *5 claims* \[6\] Yann LeCun advocates for 'world models' and the JEPA architecture as the future of AI, believing current large language models are insufficient for achieving human-level intelligence. *5 claims* \[22, 30, 46, 73\] Major open-weight LLMs released in 2025 and 2026 share a common architectural skeleton, identified as a Mixture-of-Experts model. *4 claims* \[2, 13, 15, 26, 27, 68\] Open AI ecosystems accelerate innovation while closed systems create bottlenecks. *4 claims* \[11, 22\] GRPO is a common RL optimizer for open LLMs. *3 claims* \[2, 18, 63\] In April 2026, six major labs (Google, Alibaba, Meta, Mistral, Zhipu AI, and DeepSeek) ship open-weight models that rival or surpass closed alternatives on practical workloads. *3 claims* \[5, 8, 9\] Benchmarks like HellaSwag, bare HumanEval, and naive Needle-in-a-Haystack long-context tests are considered functionally dead or saturated in 2026. *3 claims* \[54\] Mixture-of-Experts (MoE) models are prevalent, exemplified by Kimi K2, GLM-5.2, and Llama 4. *2 claims* \[8, 15, 39\] Predictions for 2027 include a growing interest in continual learning to enable models to adapt without full retraining. *1 claim* \[15\] ## 9. Entities Stability AI: \[7, 17, 53, 69, 91, 92, 93, 95, 140\] Mistral: \[2, 9, 10, 16, 17\] Hugging Face: \[2, 3, 10, 11, 12, 17, 22, 65, 73, 125, 126, 127, 146, 185, 186, 187, 188, 189, 213, 214, 215, 216\] Meta Llama: \[2, 9, 16, 17, 128\] Dolly: \[6, 17, 29, 33, 34, 35, 71, 110, 171\] Gemma: \[2, 9, 10, 16, 17, 101\] Llama 4 Scout: \[2, 11, 12, 16, 24, 40, 41, 60, 64, 84, 118, 119, 120\] Mistral Large 3: \[2, 16\] Gemma 3 27B: \[2, 16\] GLM-5.2: \[2, 20, 31, 61, 62, 81, 82, 114, 115, 183\] Kimi K2.7 Code: \[2\] DeepSeek V4 Pro: \[2, 11, 12\] MiniMax M3: \[2\] Qwen 3.6: \[2, 5\] MiniMax abab: \[2\] open-weight LLMs: \[13, 14, 26, 27, 32, 42, 43, 44, 66, 67, 133\] arXiv [cs.LG]: \[13, 26, 27\] Reflection AI: \[44\] ROCm 6: Llama2: AMD: AMD Instinct MI300A: Andreessen Horowitz: \[13, 14, 32, 42, 43, 44, 66, 67, 133\] Future of Life Institute: \[14\] Orbital: \[44\] Qualcomm: \[1\] BLOOM 176B: Yann LeCun: \[4, 10, 13, 14, 19, 22, 26, 27, 36, 37, 46, 47, 48, 49, 50, 51, 72, 73, 74, 75, 76, 77, 102, 103, 104, 105, 106, 146, 147, 148, 149, 150, 151, 152, 153, 202, 205, 206, 207\] 2026: \[13, 14, 26, 27, 32, 42, 43, 66, 67\] International AI Safety Report 2026: \[13, 26, 27\] AI Safety Index: \[14\] Mark Zuckerberg: \[14\] 2025: \[13, 26, 27\] Yoshua Bengio: \[4, 13\] AI: \[13, 14, 32, 42, 43, 66, 67\] a16z Speedrun: \[44\] AMD Instinct MI300X: \[18, 38, 52, 85, 86, 129, 130, 154, 155, 156, 157, 158, 159, 160, 161, 162, 209, 210, 211\] Qualcomm Snapdragon X Elite: \[78, 107, 108, 109, 163, 164, 165, 166, 167, 168, 169, 170\] LLM: \[10, 19, 22\] Hexagon NPU: vLLM: Azure OpenAI: \[38, 52\] Nvidia H100 HGX: CDNA 3: Zen 4: Falcon 1B: \[7\] Phi-4: \[9, 11, 12\] large language models: \[4, 50, 106\] Geoffrey Hinton: \[4\] Google DeepMind: \[4\] Llama 4: \[1, 5, 8, 17, 24, 40, 41, 60, 62, 64, 84, 118, 119, 120, 121, 128\] Qwen: \[9\] All MiniLM L6 (v2): \[65\] Google’s ELECTRA Base Discriminator: \[65\] world models: \[4, 19, 36, 46, 47, 49, 50, 72, 74, 76, 77, 102, 104, 147, 148, 149, 150, 151, 152, 203\] Turing Award: \[4, 49\] MosaicML: \[6, 29, 34, 35, 45, 98, 99, 110, 143, 144, 145, 172, 173, 174, 175, 176, 177, 178\] Gemini 1.0 Pro: \[6, 28\] GPT-4 Turbo: \[6\] Gemini: \[10, 101\] LLaMA 2: \[10, 19\] DeepSeek: \[3, 9\] AMI Labs: \[4, 19, 36, 37, 47, 48, 49, 50, 51, 74, 75, 76, 77, 153\] enterprise LLM: \[28, 33, 71\] Grok-1: \[6\] Anthropic: \[10, 19\] Mistral Large: \[9\] Gemma 2B/7B: \[9\] NSFW Image Detection: \[65\] Meta: \[4, 5, 8, 10, 17, 19, 24, 37, 40, 41, 48, 51, 60, 64, 73, 84, 103, 105, 106, 120, 121\] Advanced Machine Intelligence: \[4, 37\] Project Tapestry: \[4\] AI Alliance: \[4\] Facebook A.I. Research: \[4\] Alex LeBrun: \[4\] Nabla: \[4\] open-weight LLM: \[2, 6, 7, 8, 17, 33, 68, 69, 70, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 100, 122, 123, 124, 134, 135, 136, 137, 138, 139, 140, 141, 192, 193, 194, 195, 196, 197, 198, 199\] GPT-3.5: \[6, 28\] Mixtral: \[6\] Clément Delangue: \[22\] Databricks: \[6, 28, 29, 33, 34, 35, 45, 71, 98, 99, 100, 110, 142, 143, 144, 145, 171, 172, 173, 174, 175, 176, 177, 178, 201, 212\] DBRX: \[6, 35, 100\] MPT: \[6, 45\] CodeLLaMA-70B: \[6\] LLaMA2-70B: \[6\] Mistral Medium: \[6\] OpenAI: \[10, 19, 38, 52\] JEPA: \[19, 46\] AGI: \[104\] Stable LM 2 1.6B: \[7\] Stable Diffusion: \[69, 94, 141\] Stable Audio 3.0: \[69\] Mixture-of-Experts: \[68, 200\] European Open Digital Ecosystem Strategy: \[87\] China: \[89, 90, 96\] US: \[96\] Phi-1.5: \[7\] Phi-2: \[7\] TinyLlama 1.1B: \[7\] Open LLM Leaderboard: \[7\] ARC Challenge: \[7\] GLM-5.1: \[5, 11, 12\] DeepSeek V4 Flash: \[11\] DeepSeek R1: \[3, 12, 20, 23, 31, 58, 61, 62, 82, 83, 113, 114, 115, 116, 117, 181, 182, 183\] Moonshot AI: \[12, 20, 23, 24, 40, 41, 55, 56, 57, 58, 59, 64, 79, 80\] David Chou: \[1\] Nvidia: \[1\] Gemma 4 26B A4B: \[11\] GLM-5: \[12\] Kimi K2: \[20, 23, 55, 56, 57, 58, 59, 62, 79, 80\] Apache 2.0 license: \[1, 111\] GPT: \[1\] Kimi K2.5: \[12, 20, 23, 55, 56, 57, 58, 59, 79, 80, 113\] Google Gemini: \[112, 179\] Caren Chang: \[1\] Qwen3: \[11, 12\] RLVR: \[31, 61, 63, 83\] Z.ai: \[12, 20, 81\] Google Gemma 4: \[1, 111\] Vertex AI: \[1\] Jim Fan: \[1\] Kimi K2.6: \[11, 12, 23, 55, 56\] Gemma 4: \[5, 11, 12, 70, 97, 122, 123, 124\] GRPO: \[31, 63, 83\] Meta AI: Kimi K2 Thinking: \[20, 57\] Llama 4 Maverick: \[24, 40, 41, 60, 64, 84, 118, 119, 121\] Google: \[1, 5, 8, 111, 112, 123, 179, 217, 218, 219, 220\] Gemini Nano 4: \[1\] Google Cloud: \[1, 112\] DeepSeek V4: \[1, 5, 8\] Stable LM 2: \[7, 53\] NVIDIA A100 GPUs: \[53\] Microsoft Azure: \[38, 52\] Chat GPT: \[38\] Snapdragon X Elite: \[78, 107, 108, 109, 163, 164, 165, 166, 167, 168, 169, 170\] LM Studio: \[78\] Stability AI Community License: \[53\] Zhipu AI: \[5\] Mistral AI: \[5, 8\] Mistral Small 4: \[5\] Open Source LLM: \[5, 8, 70, 124, 225\] OpenClaw: \[59, 116\] Alibaba: \[5, 8\] Qwen 3.5: \[5\] DeepSeek-R1: \[5\] ## 10. Sources [1] *(cited 34×)* [2] *(cited 33×)* [3] *(cited 26×)* [4] *(cited 26×)* [5] *(cited 24×)* [6] *(cited 19×)* [7] *(cited 18×)* [8] *(cited 17×)* [9] *(cited 15×)* [10] *(cited 15×)* [11] *(cited 14×)* [12] *(cited 13×)* [13] *(cited 10×)* [14] *(cited 10×)* [15] *(cited 10×)* [16] *(cited 9×)* [17] *(cited 9×)* [18] *(cited 9×)* [19] *(cited 9×)* [20] *(cited 9×)* [21] *(cited 8×)* [22] *(cited 8×)* [23] *(cited 8×)* [24] *(cited 8×)* [25] *(cited 7×)* [26] *(cited 7×)* [27] *(cited 7×)* [28] *(cited 7×)* [29] *(cited 7×)* [30] *(cited 7×)* [31] *(cited 7×)* [32] *(cited 6×)* [33] *(cited 6×)* [34] *(cited 6×)* [35] *(cited 6×)* [36] *(cited 6×)* [37] *(cited 6×)* [38] *(cited 6×)* [39] *(cited 6×)* [40] *(cited 6×)* [41] *(cited 6×)* [42] *(cited 5×)* [43] *(cited 5×)* [44] *(cited 5×)* [45] *(cited 5×)* [46] *(cited 5×)* [47] *(cited 5×)* [48] *(cited 5×)* [49] *(cited 5×)* [50] *(cited 5×)* [51] *(cited 5×)* [52] *(cited 5×)* [53] *(cited 5×)* [54] *(cited 5×)* [55] *(cited 5×)* [56] *(cited 5×)* [57] *(cited 5×)* [58] *(cited 5×)* [59] *(cited 5×)* [60] *(cited 5×)* [61] *(cited 5×)* [62] *(cited 5×)* [63] *(cited 5×)* [64] *(cited 5×)* [65] *(cited 4×)* [66] *(cited 4×)* [67] *(cited 4×)* [68] *(cited 4×)* [69] *(cited 4×)* [70] *(cited 4×)* [71] *(cited 4×)* [72] *(cited 4×)* [73] *(cited 4×)* [74] *(cited 4×)* [75] *(cited 4×)* [76] *(cited 4×)* [77] *(cited 4×)* [78] *(cited 4×)* [79] *(cited 4×)* [80] *(cited 4×)* [81] *(cited 4×)* [82] *(cited 4×)* [83] *(cited 4×)* [84] *(cited 4×)* [85] *(cited 3×)* [86] *(cited 3×)* [87] *(cited 3×)* [88] *(cited 3×)* [89] *(cited 3×)* [90] *(cited 3×)* [91] *(cited 3×)* [92] *(cited 3×)* [93] *(cited 3×)* [94] *(cited 3×)* [95] *(cited 3×)* [96] *(cited 3×)* [97] *(cited 3×)* [98] *(cited 3×)* [99] *(cited 3×)* [100] *(cited 3×)* [101] *(cited 3×)* [102] *(cited 3×)* [103] *(cited 3×)* [104] *(cited 3×)* [105] *(cited 3×)* [106] *(cited 3×)* [107] *(cited 3×)* [108] *(cited 3×)* [109] *(cited 3×)* [110] *(cited 3×)* [111] *(cited 3×)* [112] *(cited 3×)* [113] *(cited 3×)* [114] *(cited 3×)* [115] *(cited 3×)* [116] *(cited 3×)* [117] *(cited 3×)* [118] *(cited 3×)* [119] *(cited 3×)* [120] *(cited 3×)* [121] *(cited 3×)* [122] *(cited 3×)* [123] *(cited 3×)* [124] *(cited 3×)* [125] *(cited 2×)* [126] *(cited 2×)* [127] *(cited 2×)* [128] *(cited 2×)* [129] *(cited 2×)* [130] *(cited 2×)* [131] *(cited 2×)* [132] *(cited 2×)* [133] *(cited 2×)* [134] *(cited 2×)* [135] *(cited 2×)* [136] *(cited 2×)* [137] *(cited 2×)* [138] *(cited 2×)* [139] *(cited 2×)* [140] *(cited 2×)* [141] *(cited 2×)* [142] *(cited 2×)* [143] *(cited 2×)* [144] *(cited 2×)* [145] *(cited 2×)* [146] *(cited 2×)* [147] *(cited 2×)* [148] *(cited 2×)* [149] *(cited 2×)* [150] *(cited 2×)* [151] *(cited 2×)* [152] *(cited 2×)* [153] *(cited 2×)* [154] *(cited 2×)* [155] *(cited 2×)* [156] *(cited 2×)* [157] *(cited 2×)* [158] *(cited 2×)* [159] *(cited 2×)* [160] *(cited 2×)* [161] *(cited 2×)* [162] *(cited 2×)* [163] *(cited 2×)* [164] *(cited 2×)* [165] *(cited 2×)* [166] *(cited 2×)* [167] *(cited 2×)* [168] *(cited 2×)* [169] *(cited 2×)* [170] *(cited 2×)* [171] *(cited 2×)* [172] *(cited 2×)* [173] *(cited 2×)* [174] *(cited 2×)* [175] *(cited 2×)* [176] *(cited 2×)* [177] *(cited 2×)* [178] *(cited 2×)* [179] *(cited 2×)* [180] *(cited 2×)* [181] *(cited 2×)* [182] *(cited 2×)* [183] *(cited 2×)* [184] [185] [186] [187] [188] [189] [190] [191] [192] [193] [194] [195] [196] [197] [198] [199] [200] [201] [202] [203] [204] [205] [206] [207] [208] [209] [210] [211] [212] [213] [214] [215] [216] [217] [218] [219] [220] [221] [222] [223] [224] [225]