1. Distilled LivePortrait Model Runs at 25 FPS in the Browser via WebGPU
A developer has released a distilled version of the LivePortrait model that runs entirely within the browser using WebGPU. While the original ONNX model required 30 seconds per frame, this distilled version achieves real-time performance of 25 frames per second (under 30ms per frame on an Nvidia 5090). Although the current output quality is limited due to a brief training run on a small dataset, it demonstrates the viability of client-side, hardware-accelerated portrait animation.
- • A developer created a distilled version of the LivePortrait model optimized for Chrome via WebGPU.
- • The original ONNX version of LivePortrait took 30 seconds to generate a single frame in the browser.
- • The distilled model generates a frame in less than 30ms on an Nvidia 5090 GPU, running entirely in-browser.
- • The model currently has limited output quality due to being trained on a small dataset for only a few hours.
Unlocks high-performance, low-latency local video and portrait animation features inside web applications without requiring expensive server-side GPU rendering.
2. SupraLabs Releases 800M Model to Summarize Agent Thought Traces
SupraLabs has launched an 800M parameter model specifically trained to parse raw agent reasoning and tool calls into structured JSON summaries. This allows developers to display clean, readable progress updates to end-users instead of raw model logs. Along with the model, SupraLabs has open-sourced a 61k-entry dataset on Hugging Face to support further development in agent trace summarization.
- • SupraLabs released the SupraLabs/reasoning-summarizer-800m-pre-gguf model to summarize agent thought traces.
- • The model processes raw reasoning and tool calls to output a structured JSON object containing a title, sub-title, summary, and current task.
- • An open-source dataset containing 61,000 reasoning summaries ('reasoning-summaries-61k') was released alongside the model on Hugging Face.
- • The model currently has a limited vocabulary for starting summaries, with a specialized version for agentic coding in development.
Enables developers to present clean, structured, and user-friendly reasoning summaries instead of exposing raw, messy agent thought traces.
3. SupraLabs Releases 51M Parameter Prompt Routing Model
SupraLabs has open-sourced Supra-Router-51M, a highly compact 51-million-parameter model designed for dynamic prompt routing. By evaluating incoming user prompts, the model determines whether to route the request to a small, fast model or a larger, more capable one. This low-latency orchestrator is available on Hugging Face alongside its training dataset.
- • SupraLabs released Supra-Router-51M, a prompt routing model with 51 million parameters.
- • The model is optimized for low-latency environments to dynamically route user requests to smaller or larger models.
- • Both the model weights and the training dataset have been made available on the Hugging Face Hub.
Helps developers reduce inference costs and latency by dynamically routing simple queries to smaller models and complex ones to larger models.
4. shadcn/ui Switches Default Component Library from Radix to Base UI
In a major architectural shift, shadcn/ui has made Base UI its default underlying component library, moving away from Radix. While Radix remains fully supported and accessible via a CLI flag, the transition reflects a strong developer preference, with new projects choosing Base UI by a 2-to-1 margin. To ease the transition, shadcn has introduced a migration skill compatible with AI coding assistants like Cursor and Claude Code to automate component-by-component refactoring.
- • As of July 2026, Base UI is the default component library for shadcn/ui, replacing Radix which was the default since launch.
- • Base UI is currently at version 1.6.0 and has over 6 million weekly downloads.
- • New projects created via shadcn/create select Base UI over Radix at a 2-to-1 ratio.
- • Radix is not deprecated and will continue to receive support and new component releases.
- • Developers can still initialize new projects with Radix using the -b radix flag during the shadcn init process.
- • A new migration skill compatible with Claude Code and Cursor is available to help transition components from Radix to Base UI.
Developers starting new shadcn/ui projects will now build on Base UI by default, though they can opt out or use automated migration tools to transition existing components.
5. Qualcomm Launches GenieX for Local LLM Execution on Windows Laptops
Qualcomm has launched GenieX, a software development kit designed to run LLMs locally on Windows laptops. The SDK integrates with llama.cpp, allowing developers to execute any Q4_0 GGUF model across Qualcomm's CPU, GPU, and NPU hardware. Early user benchmarks show promising performance, reaching 20 tokens per second on a Gemma 4 26B model using hardware acceleration.
- • Qualcomm launched GenieX to enable local LLM execution on Windows laptops.
- • The SDK allows running any Q4_0 GGUF model on Qualcomm hardware (CPU, GPU, and NPU) using llama.cpp.
- • A user reported achieving 20 tokens per second running the Gemma 4 26B A4B model on a Qualcomm GPU or NPU.
- • The same user achieved 10 tokens per second running the Qwen 3.6 27B MTP model on a Qualcomm GPU.
Enables developers to build and run local AI applications on Snapdragon-based Windows laptops, leveraging hardware acceleration.
6. Claude Design System Prompt Released to Prevent Generic AI Aesthetics
Claude Design has released an open-source, MIT-licensed system prompt and skill library aimed at improving the design output of LLMs. The system provides 20 chapters of design philosophy and 14 procedural skills to help models generate clean, accessible, and professional interfaces. By explicitly forbidding common AI design clichés like excessive emojis and heavy gradients, the prompt helps developers generate production-ready UI code.
- • Claude Design is an open-source, MIT-licensed system prompt and skill library for UI design collaboration.
- • The system includes 20 chapters of design philosophy and 14 invokable procedural skills for production, extraction, and review.
- • The prompt explicitly instructs LLMs to reject common AI design tropes like aggressive gradients, emoji decorations, and generic SaaS templates.
- • It is compatible with Claude, GPT, Gemini, and local models, with a specific variant available for OpenAI's Codex.
- • The prompt is calibrated for Anthropic's frontier models, including Fable 5 and the Opus 4.7/4.8 lineage.
Helps developers steer LLMs away from generic SaaS templates and aggressive gradients, ensuring cleaner and more accessible UI code.
7. Long-Context Benchmark Identifies Key Bottlenecks for Local Agentic Workloads
A comprehensive benchmark of 13 local LLMs running on llama.cpp highlights critical architectural factors for long-context agentic workloads. The study shows that prefill speed dominates wall-clock time, rendering standard token generation metrics secondary. Developers building local agents can leverage these findings—such as the superior scaling of 4 KV head architectures and the performance penalties of quantized KV caches—to optimize their hardware configurations.
- • The benchmark evaluated 13 local LLMs at 65K-128K context lengths using an RX 7900 XT 20GB GPU and llama.cpp build 9860.
- • Prefill speed accounts for 94% to 99% of wall-clock time for short agentic outputs at 65K context, making token generation speed largely irrelevant.
- • Models with 4 KV heads scaled significantly better for long-context prefill speed than those with 8 KV heads.
- • Mamba2 hybrid models like Granite-4.0-H-Small retained 69% of their 4K context speed with near-flat prefill scaling up to 131K context.
- • Using F16 KV cache can be 20% to 53% faster than quantized Q8 or Q4 KV cache for MoE and small dense models due to dequantization overhead.
- • Multi-head Latent Attention (MLA) architectures like GLM-4.7-Flash showed poor prefill scaling on Vulkan, losing 80% speed by 16K and crashing at 65K.
Allows developers to optimize local model selection and configuration for long-context agentic tasks without wasting VRAM or compute.
8. oMLX Optimizations Significantly Boost DeepSeek V4 Inference on Apple Silicon
Following earlier reports on running DeepSeek V4 on Apple Silicon via SSD streaming and baseline benchmarks, developers have released oMLX optimizations that substantially improve performance. By implementing native DeepSeek affine MoE kernels and 8-bit affine Metal kernel instantiations, the patches achieve a 1.6x speedup in prefill (533 tok/s) and a 3x speedup in decode (up to 22 tok/s) on a Mac Studio M3 Ultra.
- • New oMLX patches enable native DeepSeek affine MoE kernels and 8-bit affine Metal kernel instantiations.
- • Performance on Mac Studio M3 Ultra improved to 533 tok/s prefill and up to 22 tok/s decode.
- • These optimizations build on existing methods for running DeepSeek V4 on Apple Silicon, such as SSD streaming.
- • The author is currently soliciting community feedback regarding potential accuracy trade-offs.
These optimizations provide a major performance gain for developers running massive MoE models locally, moving beyond the baseline capabilities previously reported.
9. ByteDance and Alibaba Disable Custom AI Agents in China to Comply with New Rules
ByteDance and Alibaba are disabling custom AI agent features on their respective Doubao and Qwen platforms in China. This move is a direct response to the Interim Measures for the Administration of Artificial Intelligence Anthropomorphic Interaction Services, which take effect on July 15. The regulations target AI systems that simulate human personalities or foster sustained emotional interactions, though standard workplace assistants and customer service bots remain exempt.
- • ByteDance's Doubao and Alibaba's Qwen are disabling custom AI agent features in response to upcoming regulations.
- • The Interim Measures for the Administration of Artificial Intelligence Anthropomorphic Interaction Services take effect on July 15.
- • The regulations target AI services that simulate human personality, thinking patterns, and communication styles for sustained emotional interaction.
- • The rules aim to mitigate risks like extremist ideas, privacy leaks, mental health harm, and AI addiction.
- • Customer service bots, workplace assistants, and scientific research tools are excluded from the rules, provided they avoid sustained emotional interaction.
Developers shipping conversational or emotional AI agents in China must adapt to strict new compliance rules or risk service disruption.