Meituan updates LongCat Avatar to v1.5 with faster 8-step inference and INT8 support

·1 views

Meituan's LongCat team released version 1.5 of its open-source talking-avatar model on May 21, 2026, introducing several technical improvements over the December 2025 original. The update replaces the Wav2Vec2 audio encoder with Whisper-Large-v3, which the team says produces smoother lip synchronization. A new distillation flag reduces generation to just 8 sampling steps, while an INT8 quantization option lowers VRAM requirements, making the 13.6-billion-parameter model more accessible on consumer GPUs. The release also introduces Cross-Chunk Latent Stitching, a technique that eliminates redundant encoding between video segments to enable seamless long-form output. Meituan claims the model outperforms rivals including HeyGen and Kling Avatar 2.0 in internal human evaluations, though no independent benchmark has yet verified these results.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

LinkedIn's Dynamic CSS Classes Keep Breaking Third-Party Browser Extensions

A LinkedIn News blocker browser extension stopped working after LinkedIn's auto-generated CSS class names changed, severing the selectors the extension relied on to identify page elements. Modern front-end frameworks like React and Angular routinely rehash class names during build updates, even when the visible interface remains unchanged, making third-party tools fragile. The problem is compounded by potential A/B testing and regional variations, which can cause inconsistent failures across different users. Developers have identified three mitigation approaches: chasing updated selectors, using structural or text-based CSS selectors, and injecting custom data attributes via content scripts. Experts consider the last approach most robust, as it decouples the extension from LinkedIn's build process rather than depending on volatile platform-generated identifiers.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Open-Source Tool Reveals Hidden Token Waste in AI Coding Agents

A developer has released AI Agent Profiler, a local reverse proxy tool that sits between coding agents like Claude Code and LLM providers to track exactly how API tokens are spent. The tool found that only around 60% of API costs stem from actual user prompts, with the remaining 40% attributed to agent overhead such as subagent searches, context summaries, and session-title generation. A key cost driver identified is Claude Code's 5-minute cache expiry window, which forces expensive full cache rewrites on tokens exceeding 200K whenever a user steps away briefly. The profiler classifies requests into 11 categories and flags cold-cache regenerations with severity badges to help users understand idle-gap costs. The tool supports multiple providers including Anthropic, OpenAI, DeepSeek, AWS Bedrock, and Ollama, and can be installed via npm with zero telemetry.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

University of Chicago Team Demonstrates Key Circuit for Topological Quantum Computing

Researchers at the University of Chicago have built a non-planar superconducting qubit using a crossbar 'waffle grid' of three horizontal and three vertical wires, forming nine Josephson junctions. Unlike conventional flat circuit designs, this geometry enables a mathematical property called Z3 combinatorial gauge symmetry when exposed to a controlled magnetic field. Experiments confirmed the circuit settled into six equivalent low-energy states, matching theoretical predictions and providing the first lab evidence of this symmetry in engineered hardware. This symmetry is considered a foundational requirement for creating quantum spin liquids, which in turn could serve as a platform for topological quantum computing. The findings were validated using neural-network variational Monte Carlo simulations, and the work establishes a critical building block for assembling larger, error-resistant quantum systems.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

AirMeter Mod Streams Multimeter Readings Live to a Browser via ESP32

A YouTuber known as Bits und Bolts has built AirMeter, an open-source hardware mod that wirelessly streams live readings from an ANENG AN870 multimeter to a web browser or OBS. The project was created to solve the common filming problem of glare, focus hunting, and awkward camera angles when capturing multimeter screens. Because the meter's built-in UART port is physically disabled under epoxy, the mod instead uses an RP2040 microcontroller to directly sample the LCD's multiplexed analog signals, which are then converted to clean digital levels using LM339 comparators. The decoded readings are transmitted over a 433 MHz HC-12 radio link to an ESP32-C3, which hosts a local web page that faithfully recreates the meter's display. All schematics, board files, and firmware are publicly available, and the method can be adapted to other budget multimeters built on the same DTM0660 chip.

0 comments Read more at DEV Community