ExLlamav3 Gets a Major Speedup with DFlash Support, Setting a New Standard for Local LLM Inference

ExLlamav3, a local LLM inference engine, has been evolving rapidly over the past few weeks. Starting with Gemma 4 support and followed by cache efficiency improvements, developer turboderp released DFlash support two weeks ago.

DFlash brings dramatic improvements to inference speed, achieving substantial speedups across many use cases: 3.00x over the previous baseline on coding tasks, 2.51x on agentic code generation tasks, and 2.06x on translation (reasoning) tasks.

The latest version also pushes model optimization further, confirming inference speed improvements of 21.0% for Qwen3.5-35B-A3B on an RTX 5090 setup and as much as 72.4% for Trinity-Nano. For users running LLMs in local environments, the range of practical options has expanded significantly.

Google Finance's AI Features Expand to Europe

Google has started the European rollout of Google Finance equipped with new AI-powered features. Fully localized in local languages, features that let AI assist with analyzing and visualizing investment information are now available to users across European countries.

The AI-powered finance tool is designed to help retail investors understand market data more intuitively, and can be seen as part of a broader movement toward democratizing financial information.

DTap: Making Large-Scale Red Teaming of AI Agents Possible

After 20 months of development and $120,000 in API costs, the AI agent security testing platform "DecodingTrust-Agent Platform (DTap)" has been open-sourced.

DTap simulates more than 50 real-world environments across 14 high-risk domains and comes with realistic agent interfaces replicated from official MCPs and GUIs. It was released alongside "DTap-Bench," a benchmark containing roughly 7,000 red teaming tasks and about 4,000 policy-grounded malicious goals.

An evaluation of major agent frameworks and backbone models revealed that current AI agents contain systematic vulnerabilities and zero-days. It is a critical step toward ensuring agent safety.

AEM: Automatically Tuning the Exploration-Exploitation Balance in Multi-Turn RL Agents

Reinforcement learning (RL) continues to improve LLM agent performance, but one problem has remained: assigning appropriate "credit" to each step within long conversation trajectories. Prior approaches introduced dense intermediate supervision such as process reward models, which ended up adding tuning complexity.

The newly proposed "AEM (Adaptive Entropy Modulation)" is a method that performs credit assignment without supervision. It lifts entropy dynamics from the token level to the response level, automatically adjusting the balance between exploration and exploitation. Experiments on ALFWorld, WebShop, and SWE-bench-Verified confirmed consistent improvements across a wide range of model sizes from 1.5B to 32B. When integrated into a state-of-the-art software engineering RL framework, it achieved a +1.4% improvement.

TextWeb: Rendering Web Pages as Markdown for LLM Agents

A Markdown web renderer for LLM agents called "TextWeb" has been released in Reddit's LocalLLaMA community.

Conventional screenshot-based approaches depend on vision models and are costly, but TextWeb fully supports JavaScript execution while outputting web pages in Markdown format that LLMs can natively understand. It ships both a CLI and an MCP server, and can also work with llama.cpp's Web UI.

It supports navigating web pages, scrolling, text input, button clicks, and more — a tool that greatly raises the practicality of local LLM agents.

SCOPE: Tracking Complex Image Generation Instructions with Structured Specifications

Text-to-image generation models have advanced greatly in visual quality, but faithfully realizing complex user intent remains a challenge. A research team named the problem of multiple constraints becoming fragmented at each stage of the generation pipeline the "Conceptual Rift," and proposed "SCOPE" as a solution.

SCOPE represents user intent as a structured specification composed of "entities, constraints, and unknowns," and tracks it consistently across the entire generation lifecycle. It achieved 0.60 EGIP on the new benchmark "Gen-Arena" and 0.907 on WISE-V, substantially improving fidelity to complex image generation instructions.

---