2026-08-20 · ThursdayPublished daily at 08:00
AI Daily
- 1Liquid AI releases LFM2.5 series QAD Q4_0 quantized checkpoints, recovering 97% of lost accuracyHugging Face Community Blog
- 2GLM-5.3 launches: 60 on the AA Intelligence Index, tied for first among open-source models, at lower costWeChat official account: Zhipu
- 3FastMetal generates videos locally on a Mac in 30 secondsUCSD Hao AI Lab
01Model Releases/Updates
Liquid AI releases LFM2.5 series QAD Q4_0 quantized checkpoints, recovering 97% of lost accuracy
Liquid AI releases four Q4_0 GGUF checkpoints trained using quantization-aware distillation (QAD): LFM2.5-230M, 350M, 1.2B-Instruct, and 2.6B. While maintaining the memory usage and speed of native Q4_0, they recover 97% of the average BF16 precision loss.
Original text02Product Releases/Updates
GLM-5.3 launches: 60 on the AA Intelligence Index, tied for first among open-source models, at lower cost
The GLM-5.3 API launches today. It excels at complex coding, defensive cybersecurity, and long-horizon tasks, scoring 60 on the AA Intelligence Index, placing it at the same level as closed-source flagships Claude Fable 5 and GPT-5.6 Sol, and tying with Kimi K3 for first place among open-source models. With a smaller parameter scale and lower inference costs, the model lowers the threshold for frontier intelligence. Its per-task cost is the lowest among flagship models. API pricing matches GLM-5.2, and the model weights will be open-sourced next Friday.
Original textFastMetal generates videos locally on a Mac in 30 seconds
A 5-second 480P video, generated entirely on a Mac, takes 30 seconds. No CUDA or cloud is required, and it uses only 3.9 GiB of memory. FastMetal brings the FastWan-QAD series to Apple Silicon. The DiT, DMD sampler, and decoder all run on Metal through MLX, using INT8 by default. Three models are available: 1.3B supports 480P, 5B supports 720P, and 14B prioritizes image quality. 📷 Blog: http://haoailab.com/blogs/fastmetal 📷 Code: http://github.com/hao-ai-lab/FastVideo 📷 Models: http://huggingface.co/collections/Fas…
Original textGoogle Search launches 5 AI learning tools
Google launches 5 AI learning tools in Search. The generative UI for AI Mode is now available globally in English, supporting interactive visualizations and custom simulations. Practice quizzes are now available for free in English in AI Overviews and AI Mode, covering standardized tests such as the ACT and SAT.
Original textReplit launched Free Mode powered by GPT-5.6 Luna for building software at no cost
Replit introduces Free Mode, powered by GPT-5.6 Luna, allowing users to turn ideas into runnable software without worrying about token costs. Leveraging the price-performance advantages of the GPT-5.6 series and OpenAI's recent price cuts, the mode is available to millions of users and provides fast answers, suggestions, and project analysis. Users can switch to GPT-5.6 Sol when more advanced reasoning is needed while retaining project context.
Original textClaude Code v2.1.236 released: adds a default model environment variable and cross-session idle notifications
Claude Code v2.1.236 adds the ANTHROPIC_DEFAULT_MODEL environment variable, which sets the default model for new sessions. The /model selection can still override it and persists across restarts.
Original text03Industry News
Anthropic slows model development in the era of critical cyber capabilities: pauses RL training and strengthens safety protections
Original textOpenRouter announces its addition to Stripe
OpenRouter announced a merger with Stripe to accelerate global economic growth. OpenRouter currently processes more than 10 trillion tokens per day from over 400 AI models, serving more than 10 million developers and companies. Since its founding, its inference volume has grown at least 10-fold annually. After the merger, OpenRouter will continue to operate independently under its original name and mission, with no changes to its products or roadmap. Routing decisions will continue to prioritize users' interests, and the transaction is expected to close within the next few weeks.
Original text04Paper research
Breaking through the limits of DeepSeek-V4-Pro services: a multi-scenario optimization method on H20
For the 1.6-trillion-parameter MoE model DeepSeek-V4-Pro, the LMSYS team approached B300 performance on H20 GPUs through scenario-based serving configurations. A single-node H20-141GB reference implementation achieved 271 output tokens/s, narrowing the performance gap with B300 at 383.7 tokens/s to 1.42×.
Original textApple research: A multidimensional analysis of human-like behavior in LLMs: the effects of model behavior, user factors, and system prompts
Apple's machine learning research team conducted a multidimensional analysis of humanlike behavior in LLMs, such as expressing thoughts and emotions, forming relationships with users, and refusing requests while maintaining boundaries. The analysis covered the prevalence, potential impact, and controllability of such behavior. The study combined LLM-as-a-judge with human evaluation and used a sample of more than 21,000 data points, aiming to provide researchers and practitioners with a basis for deciding when and what types of humanlike behavior to address.
Original textP-completeness of inverted-index traversal: complexity assessment of Boolean-query DAGs
Modern AI agents rely on search infrastructure to perform neuro-symbolic reasoning, often compiling it into deeply nested non-monotonic Boolean queries. Standard inverted-index query evaluation strategies face severe theoretical limitations: the stateful iterator model (Document-at-a-Time) is constrained by the structure of NC^1 formula evaluation, and the worst-case query complexity explodes exponentially to O(2^|Q|) when expanding and reconverging logic.
Original text05Tips and perspectives
Beginner's guide to the GitHub Copilot app: Manage your work with the My work panel
The My work panel in the GitHub Copilot app centralizes pull requests and issues, with four default views: All, Active, Review requests, and Done. It also supports creating custom views and filters. Users can start new agent sessions directly from issues or pull requests, select multiple items in bulk to create sessions, switch between list and table views, adjust the repository scope, or create new issues.
Original textHow Slack builds human-agent teams: conversation is knowledge
Slack Chief Product Officer Jaime DeLanghe shared best practices for turning conversations into institutional knowledge and building human-agent teams. She advocates using public channels by default so agents can learn from visible conversations, and recommends connecting context from meetings, email, calendars, and other sources to reduce repetitive work. In Slack, Claude-powered agents handle productive tasks such as drafting, summarizing, and monitoring, while humans handle review, decision-making, and handoffs, creating a cycle of collaboration.
Original textHow Databricks designs effective Genie Agents from a single prompt
Databricks released a guide exploring how to design effective Genie Agents from a single prompt. The article points out that general-purpose agents often retrieve only the first relevant data table when handling queries about revenue and similar topics, while Genie Agents can locate and answer user questions more accurately through more precise prompt design. The guide is intended to help developers optimize agent behavior and improve the accuracy and relevance of query results.
Original text