Release Notes

• Multi-Theme Support: A strategically crucial theming engine, pivotal for maintaining brand consistency and enabling highly personalized user experiences. This functionality directly enhances user satisfaction by allowing for customization tailored to individual preferences, thereby increasing session duration and frequency, thereby increasing user lifetime value. The dynamic thematic structure allows us to tap into diverse market segments by tailoring the user experience and further enables advanced A/B testing and content optimization to amplify conversion rates and customer satisfaction, with an anticipated uplift in key engagement metrics. Expansion to adaptive theming and style transfer capabilities are targeted for 9.4.2, designed to enable dynamic content adaptation and personalized user experiences.
• Landing Page Builder Structure: A foundational component-based framework, accelerating content velocity and optimizing the efficiency of marketing and content distribution efforts. Enabling rapid iteration on landing pages, permitting accelerated A/B testing of value propositions, with expected improvements to conversion rates, lead generation, and overall funnel performance. This new framework facilitates highly dynamic content delivery, increasing agility and maximizing impact across all customer touchpoints.
• Complete UI Redesign: A comprehensive, data-driven UI/UX overhaul focused on optimizing the user journey, minimizing cognitive load, enhancing accessibility, and aligning the user experience with the evolution of our brand. This leads to higher conversion rates, a reduction in user churn, and an increase in feature adoption, driving up average revenue per user (ARPU). We anticipate a measurable increase in daily active users (DAU), session duration, and user engagement metrics, setting the stage for sustained growth and enhanced platform stickiness.
• Role-Based Dashboard Overhauls: Redesigned role-specific dashboards, providing personalized insights and streamlining workflows. These improvements aim to increase user efficiency and satisfaction. These enhancements will facilitate more robust user engagement, resulting in improved retention rates. Furthermore, these changes introduce new analytical capabilities and improve our ability to provide personalized recommendations, with predictive analytics capabilities to drive revenue growth.
• Flexible Header & Footer Designs: Modular header/footer system, strategically designed to empower custom branding integration. This functionality allows us to control the narrative across all user touchpoints, thereby strengthening brand recognition and improving the user journey, which results in improved user retention and overall platform engagement.
• Variable Cards Component Library: A versatile and reusable card component library, allowing for consistent, high-quality data representation, improving the efficiency of content delivery. This standardization improves the overall user experience, aiding in comprehension and maximizing engagement, which directly translates to enhanced user conversion rates.
• New Landing Pages: Dedicated landing pages for Blog and Store to optimize user experience, further enhancing brand engagement and facilitate targeted marketing efforts, with the intent to enhance conversion rates and drive revenue growth.
• Instructor Finder Page: Redesigned Instructor Finder page.
• Event Calendar Integration: Integrated event calendar.
• Dark Mode Support: System-wide dark mode implemented.
• Integrated Whiteboard System: Enhanced whiteboard system for more efficient interactions.
• Meeting Booking Wizard: Meeting booking wizard.
• Dashboard Widgets: New dashboard widgets.
• Redesigned Learning Page: Overhauled learning page.
• Advertising Modal Design: Advertising modal design improvements.
• Enhanced Video Player: Optimized video player, now prepared for integration with AI-powered features, such as automated content summarization, chapter generation, and real-time language translation, leveraging transformer-based models, enabling increased accessibility and engagement. These updates support advanced analytics on user video consumption, with the potential for personalized content recommendations, thereby improving user retention and facilitating cross-sell opportunities.
• Flexible Card Components & Empty State Views: Refined card and empty state views for personalized content recommendations and proactive assistance, improving user satisfaction. This enhancement improves the user experience and promotes increased conversion rates.
• New Modal and Toast Designs: Introduction of new modal and toast notification designs.
• New Icon Pack and Icon Picker: Introduction of a comprehensive icon pack.
• New Contact Us Page: Enhanced contact us page.
• Creation Wizards: Introduction of new creation wizards.
• AJAX Search: Improved AJAX search.
• Redesigned Search Result Page: Enhanced search result page incorporating semantic search capabilities to aid in user content discovery.
• Quiz Enhancements: Quiz enhancements.
• Content Sharing Prevention: YouTube content sharing prevention.
• Caching Support: Caching support.

Gemma3 VLM Support - Powering the Future of AI-Driven Experiences

Comprehensive Vision-Language Integration: Seamless integration with Gemma3 VLM, driving advancements in multimodal AI, and enabling cutting-edge capabilities for our platform. These enhancements unlock new levels of user engagement, drive down costs, and revolutionize content creation. This integration supports the development of multimodal AI applications, including advanced visual search and the capability to understand image context, for enhanced and optimized content generation. We are at the forefront of the AI revolution, transforming how users interact.

• EP (Execution Plan) Support: Addition of large-scale EP support enables highly optimized model execution, critical for enhancing performance and resource utilization. This facilitates the implementation of advanced model parallelism techniques. By reducing inference latency and increasing throughput, we significantly decrease operational costs while improving our ability to deliver superior performance.
• NIXL Integration: Integrated NIXL (Neural Information eXchange Layer) into the communication layer of the disaggregated service, optimizing data transfer and decreasing end-to-end latency. These efficiencies are essential for high-throughput data movement, which supports enhanced model performance.
• Fabric Memory Support: Enabling Fabric Memory support for KV Cache Transfer. These enhancements improve performance, while minimizing hardware costs. This enhances our resource utilization and enables more efficient data transfers.
• MCP in ScaffoldingLLM: Introduced MCP (Model Compilation Pipeline) within ScaffoldingLLM to accelerate model compilation, shortening the iteration cycle. This is expected to improve time-to-market for new products and features.
• FP8 Quantization Support: Support for FP8 quantization, enabling model compression and improved efficiency, reducing resource consumption and improving performance-per-watt. With this functionality, we increase model efficiency and enable faster inference and reduced memory consumption.
• TRTLLM Sampler Optimizations: Generation logit and log probs for more advanced model capabilities. This contributes to enhanced model performance.
• Disaggregated Serving Enhancements: Enabled Disaggregated serving for Qwen-3 and EAGLE3 support. The Disaggregated serving architecture allows for optimized resource allocation, increasing performance, and enhancing our scalability.
• MoE Optimization: Enhancements to the MoE module improve performance.
• Chunked Attention: Added chunked attention to support the Blackwell and Hopper architectures, providing support for increased data scale.
• Sliding-Window Attention Kernels: Introduced sliding-window attention kernels.
• DeepSeek FP8 Optimization: Updated DeepSeek FP8 TRT-LLM Gen cubins, providing better performance for large batches.
• FP8 GEMM on SM89: Optimized for the SM89.
• Overlap Scheduler Optimizations: Enabled overlap scheduler.
• MLA Integration: MLA is integrated with piecewise CUDA.
• Fusion Optimizations: Fused functionalities to optimize the system.
• FP8 Block Scale MoE Integration: Integrated to streamline the deployment process.
• Speculative Decoding: Added support for speculative decoding, which allows for faster text generation, and allows for significantly improved throughput and performance.
• Model Validation: Validated Llama 3.1 models on H200 NVL.
• External Chatbot Enhancements: Improved connection criteria for AI + Human Agent and updated UI elements in the External Chatbot.
• AI Social Media Suite: AI Social Media Suite now included in the Pricing Plan.
• Default Chat Screen for External Chatbot: Set the Chat menu screen as the default, thus increasing user engagement.

• Benchmark Script Enhancements: Added all_reduce.py, establishing a robust baseline for measuring improvements and validating performance gains. These enhancements are crucial for quantifying the performance of our LLMs. The benchmarks offer detailed insight into performance optimization, which translates into measurable ROI.
• TRTLLM-Bench Updates: Updates to the trtllm-bench command that will help in measuring and quantifying latency. The inclusion of new parameters further enhances our ability to measure the performance of our LLMs.
• LoRA Performance Testing: Enabled trtllm-bench for LoRA and end-to-end performance testing, which is critical for enabling efficient fine-tuning of LLMs and enhancing model efficiency.
• Post-Processing Support: Added post_proc support in bench.
• Streaming and KV Cache Reuse Benchmarking: Added no_kv_cache_reuse and streaming support in bench. This allows us to optimize our infrastructure to provide top tier performance.

• Docker Base Images: Updated to the newest images, ensuring that our infrastructure remains compliant with industry standards.
• Library Upgrades: Updating to the latest libraries. This is key to our growth and performance.

• Config Object Definition: Standardized configuration object definition, simplifying the deployment of LLMs.
• Decoder Interface Refinement: Improved code modularity.
• Torch Compile Configuration: Enhanced to maximize inference speed.
• Redundancy Removal: Removed code redundancy.
• API Access Refinement: Improved API Access.

• Fixed disaggregated service hang with MNNVL two-shot AllReduce (#4678).
• Fixed EP load balancer misrouting (#4767).
• Fixed cuda graph padding for spec decoding (#4853).
• Fixed LLaMA 4 long context issue (#4809).
• Fixed max_num_sequences calculation with overlap scheduling (#4532).
• Fixed chunked prefill + overlap scheduling (#5761).
• Fixed trtllm-bench hang due to LLM API IPC (#4798).
• Fixed index out of bounds in spec decoding (#5954).
• Fixed MTP illegal memory access in cuda graph warmup (#5947).
• Fixed slot allocation error in spec decode + disagg (#5975).
• Fixed attention window one-off bug for Gemma3 1B (#5564).
• TikTok connection issue fixed
• App performance issue fixed
• Multi-model selection not visible on mobile for all themes fixed
• AI Chat Pro UI issue fixed
• GPT-5 responses show unnecessary numbers fixed
• GPT responds with the same response twice fixed
• AI Chat Pro Image Generation shows unknown numbers fixed
• Voice Bot conversations not scoped to individual users fixed
• AI Chat displays 2 models instead of one fixed
• External Chatbot message alignment issue fixed

• accuracy/test_cli_flow::TestGpt2::test_beam_search_large is broken. Investigating and will resolve in 9.4.2.
• Disaggregated serving + MTP + overlap scheduler together can cause accuracy issues. Actively working on a solution, which we expect to release in 9.4.2.
• LLaMA4 functional support (>8K seq length) introduces a performance regression on Hopper. Addressing the performance regression on Hopper architecture will be prioritized, and optimized kernels will be created for the next release.