Open Weights

Zhipu AI's latest release, GLM-5.2, has officially claimed the top spot among open-weights models on the prestigious Artificial Analysis Intelligence Index, outperforming industry stalwarts like Llama 3.1 and Qwen 2.5. ▶ A New Performance Ceiling: GLM-5.2 demonstrates exceptional proficiency in complex reasoning, code generation, and multi-turn dialogue, signaling that Chinese open-source models have fully entered the global premier league of LLM performance. ▶ Strategic Ecosystem Shift: This achievement is more than a leaderboard win; it represents Zhipu AI’s aggressive push to capture global developer mindshare through high-performance open weights, directly challenging Meta’s dominance in the open-source landscape. Bagua Insight The rise of GLM-5.2 to the top of the Artificial Analysis Index is a landmark moment for the democratization of frontier-level intelligence. Artificial Analysis is widely regarded for its rigorous, real-world benchmarking. GLM-5.2’s success highlights a critical narrowing of the "intelligence gap" between proprietary giants (like GPT-4o and Claude 3.5) and open-weights models. We are witnessing a pivot where the trade-off between private hosting and peak performance is becoming negligible. Zhipu’s rapid iteration cycle reflects the "China speed" in AI development, forcing global competitors to accelerate their release schedules or risk losing the developer ecosystem to more accessible, high-performing alternatives. Actionable Advice Enterprise architects should prioritize GLM-5.2 for pilot testing in RAG and Agentic workflows, particularly where data sovereignty and fine-tuning flexibility are paramount. Developers should monitor integration updates in inference engines like vLLM and Ollama to leverage GLM-5.2’s superior reasoning-to-latency ratio for cost-effective rapid prototyping.

Event CoreZ.ai, the prominent Chinese AI powerhouse, has officially open-sourced GLM-5.2 as of June 16. This massive 753B parameter model utilizes a Mixture-of-Experts (MoE) architecture with 40 active parameters. Released under the highly permissive MIT license, GLM-5.2 positions itself as arguably the most powerful text-only open-weights model available to the global developer community today.▶ License Aggression: By opting for the MIT license over restrictive community licenses, Z.ai is making a strategic play for ecosystem dominance, lowering the barrier for commercial integration.▶ Architectural Scale: The 753B MoE configuration balances brute-force capacity with computational efficiency, targeting the performance-to-cost sweet spot for high-end inference.▶ Textual Purity: Decoupled from the vision series, GLM-5.2 doubles down on core linguistic reasoning and complex instruction following, directly challenging the Llama 3 hegemony.Bagua InsightThe release of GLM-5.2 is more than just a performance milestone; it is a tactical strike against the licensing moats built by Meta and other Western labs. While the industry has been trending toward multimodal "everything models," Z.ai’s decision to refine a pure-text powerhouse suggests a focus on the "Reasoning" bottleneck that still plagues GenAI. The 753B scale indicates that the Scaling Law is still the primary weapon in the LLM arms race, but the MoE efficiency suggests a maturing approach to infrastructure management. By offering an MIT-licensed alternative at this scale, Z.ai is effectively "commoditizing the complement," making high-end reasoning accessible and forcing competitors to reconsider their restrictive distribution models.Actionable AdviceEnterprises specializing in high-stakes sectors like legal, finance, or complex coding should prioritize evaluating GLM-5.2 for local deployment. The MIT license provides a unique legal runway to build proprietary layers without the "Llama-style" usage constraints. Developers should assess the hardware requirements for the 40 active parameters to optimize throughput, as this model represents the new ceiling for what can be achieved with open-weights in specialized text-processing pipelines.

Core Summary Zhipu AI's GLM-5.2 has ascended to the top of the Artificial Analysis open weights leaderboard, surpassing industry stalwarts like Meta's Llama 3.1 and setting a new performance standard for Chinese-developed AI. Bagua Insight ▶ Paradigm Shift: GLM-5.2’s dominance is not merely a result of parameter scaling; it reflects superior architectural optimization in multimodal reasoning and long-context efficiency, validating the global viability of the MoE (Mixture-of-Experts) path. ▶ The Open-Source Power Struggle: This milestone shatters the Silicon Valley monopoly on open-weights performance metrics, forcing global developers to re-evaluate the reliability and performance ceiling of the Chinese AI stack for enterprise-grade applications. Actionable Advice For Engineering Teams: Conduct immediate stress tests on GLM-5.2, specifically benchmarking inference costs and accuracy against your current production models in vertical domains like legal, finance, or software engineering. For Strategic Planning: For enterprises balancing the need for sovereignty with high-performance requirements, GLM-5.2 serves as a compelling alternative. Integrate it into your multi-model deployment strategy to mitigate vendor lock-in risks.

Zhipu AI has officially transitioned GLM 5.2 from a restricted preview to a full-scale public release, offering API access, MIT-licensed weights on HuggingFace, and immediate integration within the Ollama ecosystem. ▶ Frictionless Deployment: The rapid pivot from the gated "GLM Coding" program to day-zero Ollama support removes all barriers to entry, enabling instant local integration for the global developer community. ▶ Strategic Permissiveness: By opting for the MIT license, Zhipu is positioning GLM 5.2 as a high-performance, low-friction alternative for commercial applications, directly challenging the dominance of Llama and DeepSeek in the open-weight arena. Bagua Insight The swift democratization of GLM 5.2 signals a strategic recalibration in the post-DeepSeek landscape. In today's market, "accessibility" is the new competitive moat. Zhipu is leveraging the Ollama ecosystem to bypass traditional distribution hurdles, ensuring that GLM 5.2 becomes a daily driver for the LocalLLaMA community rather than just another benchmark entry. The choice of the MIT license is a calculated move to win over enterprise users who are increasingly wary of the restrictive licensing terms found in other "open" models. It’s a classic play for ecosystem dominance: lower the floor to raise the ceiling. Actionable Advice Local-first developers should prioritize benchmarking GLM 5.2 via Ollama for coding and reasoning tasks immediately. For enterprise architects, the MIT license presents a low-risk pathway to integrate a top-tier Chinese LLM into internal RAG pipelines. It is highly recommended to evaluate GLM 5.2 as a cost-effective, compliant alternative for private cloud deployments where licensing overhead and data sovereignty are paramount.

Zhipu AI's GLM-5.2 has achieved a historic milestone by becoming the first open-weights model to surpass the 80% mark on the Terminal-Bench benchmark, outperforming all existing open-source rivals and eclipsing proprietary giants like Google Gemini in technical reasoning tasks. ▶ Open-Source Parity Achieved: GLM-5.2 represents a paradigm shift in command-line reasoning and tool-use accuracy, proving that open-weights models can match or exceed the reasoning depth of elite closed-source systems. ▶ The New Gold Standard for Agents: By delivering frontier-level performance at a fraction of the cost, GLM-5.2 is positioned as the definitive engine for the next generation of autonomous AI agents and developer tools. Bagua Insight The significance of GLM-5.2’s performance on Terminal-Bench cannot be overstated. Unlike generic benchmarks, Terminal-Bench tests a model's ability to navigate real-world CLI environments, requiring precise logic and robust error handling. GLM-5.2’s dominance suggests that Zhipu AI has cracked the code on high-density reasoning within an open-weights framework. This is a "Sputnik moment" for the open-source community; it signals that the gap between proprietary "black boxes" and transparent, deployable weights is effectively closed for technical workflows. We are moving from an era of "open-source as a backup" to "open-source as the primary choice" for mission-critical agentic infrastructure. Actionable Advice 1. For Developers: Integrate GLM-5.2 immediately into agentic workflows like Cline or Aider. Its superior terminal reasoning reduces the "trial-and-error" cycles in automated coding and system administration. 2. For Enterprise Architects: Re-evaluate your reliance on high-cost proprietary APIs for internal dev-ops tools. GLM-5.2 offers a path to SOTA-level automation with the benefits of local deployment, data sovereignty, and significantly lower inference overhead. 3. Strategic Monitoring: Watch for GLM-5.2’s integration into broader ecosystem tools. Its success on Terminal-Bench indicates a specialized optimization that could soon disrupt the market for automated software engineering (SWE) agents.

ZONOS2 is a cutting-edge real-time Text-to-Speech (TTS) model featuring an 8B total/900M active parameter architecture. It currently holds the top position on the TTSDS prosody benchmark with a score of 88.7, outperforming major incumbents. The model weights, inference, and evaluation code are now fully open-sourced. ▶ Prosody as the New Frontier: By outclassing Qwen 3 TTS and Cartesia Sonic 3.5, ZONOS2 signals a shift in industry focus from mere intelligibility to high-fidelity emotional nuance and natural cadence. ▶ Sparse Activation Efficiency: The 900M active parameter design allows ZONOS2 to deliver the reasoning depth of an 8B model while maintaining the low-latency requirements necessary for production-grade real-time applications. Bagua Insight ZONOS2 represents a significant tactical strike by the open-source community against proprietary TTS titans like ElevenLabs and Cartesia. For too long, high-fidelity, zero-shot voice cloning was gated behind expensive APIs. ZONOS2’s dominance on the TTSDS leaderboard proves that open-weights models can achieve "human-like" prosody—capturing the subtle breaths and emotional inflections that define natural speech. This release is a massive win for the LocalLLaMA ecosystem, providing the essential "voice" for local-first AI agents that require both privacy and performance. Actionable Advice Developers should prioritize benchmarking ZONOS2’s zero-shot cloning capabilities within specific vertical domains, such as gaming or interactive storytelling, where emotional range is critical. Enterprises currently reliant on costly TTS SaaS should explore ZONOS2 as a high-performance alternative to reduce OpEx while maintaining data sovereignty. We recommend optimizing the inference stack specifically for the 900M active parameter path to achieve sub-100ms TTFT (Time To First Token) in voice-first interfaces.

Event CoreZhipu AI is set to debut its latest large language model, GLM-5.2, next week. In a major strategic shift, the model will feature open weights under the highly permissive MIT license, signaling a radical commitment to transparency and global developer adoption.▶ The MIT License Pivot: Moving to an MIT license is a "nuclear option" in the open-weights space. By allowing unrestricted commercial use and derivative works, Zhipu is effectively removing the licensing friction that often plagues enterprise adoption of proprietary-grade models.▶ Aggressive Iteration Cycles: The leap to version 5.2 suggests significant architectural refinements, likely targeting SOTA performance in reasoning, long-context handling, and instruction following.Bagua InsightThis isn't just a model drop; it's a calculated play for "Developer Sovereignty." As the competition between Meta’s Llama ecosystem and proprietary giants like OpenAI intensifies, Zhipu is positioning itself as the most "freedom-centric" alternative. By adopting the MIT license, Zhipu aims to become the default engine for the next wave of RAG and Agentic workflows. This move bypasses the restrictive clauses found in Meta's acceptable use policies, offering a truly "no-strings-attached" foundation for global startups. In the high-stakes game of GenAI, Zhipu is betting that radical openness will generate the network effects necessary to sustain a global AI ecosystem despite geopolitical headwinds.Actionable AdviceEngineering leads should prepare benchmarking pipelines to evaluate GLM-5.2’s performance against Llama 3.1/4. Given the MIT license, this model is a prime candidate for deep fine-tuning and integration into proprietary software stacks where IP ownership is a non-negotiable requirement.

Chinese AI unicorn MiniMax is reportedly set to release the open weights for its flagship M3 model this Friday, a strategic pivot aimed at capturing the global developer ecosystem and challenging the dominance of established open-source giants. ▶ Competitive Benchmarking: M3’s prowess in long-context retrieval and complex reasoning positions it as a formidable challenger to Meta’s Llama 3.1 and Alibaba’s Qwen 2.5, potentially shifting the SOTA (State-of-the-Art) landscape for open-weight models. ▶ Strategic Pivot: By embracing open weights, MiniMax is transitioning from a closed-API silo to a dual-track strategy, leveraging community-driven optimization to refine its proprietary stack and reduce inference overhead. Bagua Insight The decision to open-source M3 signals a "DeepSeek moment" for MiniMax. Historically known for its high-performing closed models, MiniMax has struggled with developer mindshare compared to the aggressive open-source pushes from Alibaba and DeepSeek. Releasing M3 weights is a calculated move to gain global legitimacy. For the Silicon Valley ecosystem, this adds another high-quality Chinese model to the toolkit, further commoditizing intelligence. The real value of M3 lies in its sophisticated handling of long-context windows—a traditional pain point for open-source models—which could make it the new gold standard for local RAG (Retrieval-Augmented Generation) implementations. Actionable Advice Benchmark Immediately: Engineering teams should prioritize benchmarking M3 against Llama 3.1 for long-context needle-in-a-haystack tests and logical reasoning tasks upon release. Infrastructure Readiness: Ensure local inference environments (e.g., vLLM, TGI) are ready for testing. Monitor for GGUF/EXL2 quantizations to assess deployment feasibility on consumer-grade hardware. Monitor Fine-tuning Potential: Keep a close watch on the model's license terms. If permissive, M3 could become a superior base for domain-specific fine-tuning in sectors like legal, finance, and technical documentation.

Core Event Google has officially launched Gemma 4 12B, its first unified, native multimodal open-weights model featuring a groundbreaking "encoder-free" architecture. By moving away from external vision or audio encoders, Gemma 4 processes text, images, audio, and video within a single Transformer backbone, signaling a major paradigm shift from modular "Frankenstein" models to true multimodal integration. ▶ Architectural Revolution: By ditching external encoders like CLIP, Google eliminates information bottlenecks and synchronization issues, achieving seamless native cross-modal reasoning. ▶ Efficiency at Scale: At 12B parameters, the model delivers performance in multimodal understanding and reasoning that rivals or exceeds significantly larger proprietary models. ▶ Ecosystem Play: Google is leveraging this release to challenge Meta’s Llama dominance in the open-weights space, setting a new technical benchmark for lightweight multimodal AI. Bagua Insight Gemma 4 is more than just a performance bump; it’s a strategic pivot in AI infrastructure. For years, the industry relied on "stitching" separate encoders to LLMs, which often resulted in a loss of nuance during cross-modal translation. Gemma 4 proves that a single neural fabric can master multiple sensory inputs natively. This unified approach drastically reduces inference latency and memory footprint, making it a game-changer for on-device AI. Google is effectively democratizing the sophisticated multimodal capabilities of Gemini, signaling that the future of GenAI lies in architectural elegance rather than just brute-force scaling. Actionable Advice 1. Pivot from Modular to Unified: Developers should begin transitioning from legacy CLIP+LLM pipelines to unified architectures like Gemma 4 to reduce system complexity and technical debt. 2. Prioritize Edge Deployment: The 12B parameter count is the "sweet spot" for high-end edge devices. Organizations should explore real-time multimodal agents in sectors like automotive, robotics, and premium mobile apps. 3. Refine Multimodal Data Pipelines: Since native models thrive on interleaved data, data engineering teams should focus on curating datasets where text, audio, and visuals are deeply synchronized, rather than training on isolated modalities.

Google DeepMind has officially unveiled the Gemma 4 series, featuring a 12B multimodal powerhouse that integrates text, image, and native audio processing. With a massive 256K context window and support for 140+ languages, Gemma 4 sets a new high-water mark for open-weight efficiency and versatility. ▶ Modality Parity: Bringing native audio and vision to a 12B parameter footprint marks a strategic shift where "small" models no longer compromise on sensory input, enabling true omni-modal edge applications. ▶ Contextual Dominance: The 256K context window positions Gemma 4 as the premier choice for long-form RAG and complex enterprise document intelligence, challenging much larger proprietary models. Bagua Insight Google is executing an "asymmetric flanking maneuver" against Meta’s Llama dominance. While the industry has been fixated on scaling laws for text, Google is pivoting toward "Modality Density." By baking native audio support into the 12B class, they are targeting the next generation of voice-first AI agents and localized multimodal processing. This isn't just an incremental update; it’s a bid to capture the "Global Edge" market. Supporting 140+ languages out of the box suggests Google is prioritizing international developer adoption to build a moat that raw English-centric benchmarks cannot easily breach. Actionable Advice Engineering teams should prioritize benchmarking Gemma 4 for unified multimodal workflows to eliminate the operational overhead of managing separate models for speech, vision, and text. For RAG architectures, focus on stress-testing the 256K window's retrieval fidelity; if the "lost in the middle" effect is minimized, it could significantly simplify data ingestion pipelines by reducing the need for aggressive chunking and complex vector database strategies.

Executive Summary G7 nations have established a unified terminology framework to distinguish between "Open Source AI" and "Open Weights AI." This consensus represents a pivotal shift in global AI governance, moving from industry-led discourse to standardized international policy. ▶ Granular Regulation: By decoupling "Open Weights" from the strict OSI definition of "Open Source," the G7 is closing the loophole used by major labs (e.g., Meta) to claim open-source status while maintaining proprietary control over training data and pipelines. ▶ Foundation for Compliance: This shared language is the precursor to international enforcement mechanisms, including export controls and safety mandates, ensuring that "openness" does not become a shield against liability. Bagua Insight This is far more than a semantic exercise; it is a strategic pivot in AI geopolitics. For the past two years, the industry has operated in a "gray zone" where models like Llama enjoyed the marketing halo of open source without meeting its transparency requirements. By formalizing these definitions, the G7 is effectively narrowing the maneuver room for Big Tech. We expect this to lead to a bifurcation in regulation: "True Open Source" may receive R&D incentives, while "Open Weights" models will likely face rigorous safety audits and data provenance requirements similar to proprietary models. The G7 is signaling that the era of "Open-Washing" is officially over. Actionable Advice 1. Audit Tech Stacks: Enterprises should immediately identify dependencies on "Open Weights" vs. "True Open Source" models to anticipate shifting compliance costs in cross-border deployments. 2. Refine Procurement Standards: Update AI procurement policies to require specific disclosures on model training data and license types, as "Open Weights" models may soon carry higher insurance premiums or liability risks. 3. Monitor Policy Cascades: Watch for localized legislative updates in the UK and EU that will use these G7 definitions to trigger specific safety testing mandates for high-compute models.

Event Core Meta’s legal department has officially issued a legal notice (likely a Cease and Desist) to the creator of the Heretic project. This move, targeting a tool within the LocalLLaMA ecosystem, centers on alleged violations of Meta’s Llama Community License and trademark policies, signaling a shift in how the tech giant polices its "Open Weights" territory. ▶ Trademark Enforcement: Meta is aggressively asserting control over the "Llama" brand, targeting any project that risks brand dilution or implies an unsanctioned official endorsement. ▶ The "Open" Paradox: This incident underscores that Llama is not "Open Source" by OSI standards; it is a proprietary asset under a restrictive license that Meta is now weaponizing to prune its ecosystem. ▶ Strategic Pivot: The legal pressure on Heretic suggests Meta is moving from a phase of rapid ecosystem seeding to one of strict regulatory and brand consolidation. Bagua Insight Meta’s strategy with Llama has always been a tactical moat-building exercise rather than pure altruism. By serving Heretic, Meta is drawing a hard line in the sand: you can build on Llama, but you cannot build over it or around its branding. This is a classic Big Tech maneuver—subsidize the ecosystem with "free" tech to kill competition, then enforce strict governance once the industry is hooked. For the decentralized AI community, this is a wake-up call. The "Open Weights" movement remains fragile and beholden to the legal whims of Menlo Park. Heretic is likely just the first of many projects to be "rationalized" as Meta seeks to sanitize the Llama ecosystem for enterprise-grade optics. Actionable Advice 1. Adopt "Clean Room" Naming: Developers should pivot away from using "Llama" as a prefix or suffix. Use vendor-neutral branding and relegate model compatibility to the technical documentation to mitigate trademark infringement risks.2. License Due Diligence: Any startup leveraging Llama weights must conduct a rigorous legal audit of their distribution mechanisms, especially if they involve modified weights or bypass Meta’s standard access gates.3. Hedge with True Open Source: To avoid platform risk, maintain architectural flexibility to swap Llama for truly open models (e.g., Mistral or Apache 2.0 licensed models) should Meta further tighten the screws on its community license.

Event Core The Alibaba Qwen team has unveiled a preview of its next-generation flagship model, Qwen 3.7. This is far more than a routine version bump; it signals the formal entry of Chinese Large Language Models (LLMs) into a new epoch defined by 'Deep Reasoning' and 'Native Long Context.' Qwen 3.7 aims to achieve a quantum leap in mathematics, coding, and complex logical reasoning by implementing a 'thinking' mechanism (System 2 Reasoning) akin to OpenAI’s o1 series, all while reinforcing its dominance in the open-weight ecosystem. In-depth Details Technical disclosures indicate that Qwen 3.7’s evolution is anchored in three dimensions. First is Reinforcement Learning (RL)-driven reasoning chains: the model has transitioned from simple next-token prediction to an internal Chain-of-Thought (CoT) process that enables self-verification and path correction, drastically reducing logical hallucinations. Second is Native Support for Ultra-Long Context, with preview benchmarks showing stable processing power exceeding 1M tokens and near-perfect recall in 'Needle In A Haystack' tests. Third is the Refinement of the Mixture-of-Experts (MoE) Architecture, which significantly boosts inference efficiency per unit of compute while maintaining activated parameter scales at 32B or 72B. Commercially, Alibaba is pursuing a 'Full-Stack' release strategy, spanning from lightweight edge-side models to high-performance cloud variants. Notably, the team highlighted the Qwen-3.7-Coder variant, whose performance on benchmarks like HumanEval is now neck-and-neck with Claude 3.5 Sonnet, suggesting a lower barrier to entry for sophisticated AI Agents. Bagua Insight From a global 'Bagua Intelligence' perspective, Qwen 3.7 is reshaping the balance of power in the AI sector. While Silicon Valley has long held a first-mover advantage in 'Deep Reasoning,' Qwen is closing the gap through extreme engineering prowess and superior synthetic data utilization. For the global developer community, Qwen 3.7 provides a formidable 'Open-Weight Alternative' to closed-source giants, directly challenging the pricing power of OpenAI and Anthropic. More profoundly, Qwen 3.7 proves that even under compute constraints, exponential gains in model capability are achievable through algorithmic optimization—specifically via RL and high-fidelity synthetic data. This serves as a survival blueprint for non-US AI players. Furthermore, Qwen’s ambition in multimodal integration suggests it is aiming to set new industry standards at the intersection of visual perception and logical deduction. Strategic Recommendations For Developers: Evaluate the Qwen 3.7 Reasoning API immediately. Given its cost-performance ratio in complex logic tasks, consider migrating back-end logic from GPT-4o to Qwen to reduce operational overhead by 30%-50%. For Enterprise Leaders: Focus on the private deployment potential of Qwen 3.7. For industries like finance and law, which require deep logical analysis and have high data privacy requirements, Qwen 3.7 is currently the most viable base model. For Infrastructure Providers: The MoE architecture of Qwen 3.7 demands higher inference VRAM. Optimization of High Bandwidth Memory (HBM) allocation strategies will be critical to support the upcoming surge in long-context reasoning workloads.

Event Core The latest iteration of Alibaba's Qwen3.6 35b-a3b model has emerged as a top-tier performer in local deployment, demonstrating superior inference speed and instruction-following capabilities compared to the Gemma4 26b-a4b when executed via llama.cpp. Bagua Insight ▶ Generational Leap in Inference Efficiency: While initial performance on Ollama may vary due to abstraction overhead, the model's native execution on llama.cpp highlights significant breakthroughs in compute scheduling and MoE (Mixture-of-Experts) optimization. ▶ The Dividend of Deterministic Instruction Following: The model’s enhanced stability in complex prompting scenarios indicates that open-weights models are rapidly closing the gap with proprietary systems in production-grade reliability. Actionable Advice For developers prioritizing raw inference throughput, bypass high-level abstractions and interface directly with the llama.cpp core to fully leverage the model's hardware-level optimizations. Consider Qwen3.6 35b-a3b as the primary candidate for benchmarking RAG pipelines or complex reasoning tasks within the 30B parameter class.

Event Core The anonymous text-to-image model 'Peanut' has debuted at 8th place on the Artificial Analysis leaderboard, signaling a potential shift in the open-weights landscape as it prepares to challenge incumbents like FLUX.2 [dev]. Bagua Insight The 'Black Box' Disruption: The sudden emergence of Peanut underscores a shift in GenAI development, where anonymous contributors are now delivering performance that rivals well-funded labs. This suggests that the barrier to entry for high-fidelity image synthesis is collapsing. Beyond Parameter Scaling: Peanut’s high ranking in blind tests indicates superior prompt adherence and aesthetic coherence, suggesting that the model likely employs advanced distillation or novel architectural optimizations rather than just sheer compute power. Actionable Advice For Developers: Monitor the Hugging Face repository for the weight release. Prioritize benchmarking Peanut against your current production models, specifically focusing on VRAM efficiency and inference latency. For Enterprise Leaders: Evaluate the potential for cost-arbitrage. If Peanut proves to be a high-performance, low-latency alternative to proprietary APIs, it could significantly reduce operational overhead for image-heavy product pipelines.