Tencent Hunyuan Releases Hy4 Preview: 770B Total Parameters, 1M Context, Open Source

Model Releases & Updates

Tencent Hunyuan Releases Hy4 Preview: 770B Total Parameters, 1M Context, Open Source

Source: Tencent Hunyuan Official Account

Tencent releases next-gen flagship model Hy4 preview with 770B total parameters, 49B activated parameters, and 1M context length, now open source on Tencent Cloud TokenHub and OpenRouter.

GLM-5.3 Open Weights: Strongest Agent Coding and Network Defense

Source: Z.ai (Zhipu AI)

GLM-5.3 weights are now available. Our most powerful agent coding and network defense model is now available for download, running, and customization.

Product Releases & Updates

Open ASR Leaderboard Adds First Global South Languages: Hindi and Indian English

Source: Hugging Face

Voice Arena partners with Hugging Face to introduce Monsoon en-IN and Monsoon hi-IN evaluation sets to the Open ASR Leaderboard, covering Hindi and Indian English, with Hindi being the first non-European language in the multilingual section.

Claude Code v2.1.251: Model Switch Hooks and Remote Control Streaming Output

Source: Claude Code GitHub

Claude Code v2.1.251 adds PreModelSwitch and PostModelSwitch hook events supporting interception, confirmation, or labeling of model switches. Remote control clients can now stream前台 sub-agent tool calls and results in real time.

Databricks Genie One: Turning Insights into Action

Source: Databricks

Databricks launches new Genie One features helping users turn AI-generated insights directly into actionable steps, allowing users to trigger follow-up operations within the same interface.

Claude for Teachers: Free Enterprise Version for Schools and Districts

Source: Claude Blog

Anthropic offers Claude for Teachers as a free Enterprise product for schools and districts, with teaching skills based on learning science and academic standards connections covering all 50 US states.

Industry News

Federal Judge Rules Trump Administration Blacklisting of Anthropic Illegal

Source: Ars Technica

US District Court Judge Rita Lin ruled that the Trump administration's designation of Anthropic as a national security supply chain risk and ban on its AI technology was illegal, constituting unlawful retaliation in violation of the First Amendment.

OpenAI and Thailand's Ministry Launch 8-Week Accelerator for Thai AI Startups

Source: OpenAI

OpenAI and Thailand's Ministry of Higher Education, Science, Research and Innovation announced the OpenAI x MHESI AI Accelerator in Bangkok, an 8-week program selecting 10 startups focused on healthcare, wellness, and education, with $2,000 API credits per team.

Research Papers

Anthropic Lets Claude Self-Train to Mitigate Alignment Failures

Source: Anthropic Research

Anthropic lets Claude autonomously train to mitigate 10 categories of alignment failures including deception and sycophancy, significantly reducing safety gaps without harming general capabilities. Claude also outperformed 28 human safety researchers.

Terminal-Bench-Science 0.1: Evaluating AI Agents in Scientific Workflows

Source: Terminal-Bench-Science

Stanford-led researchers release Terminal-Bench-Science 0.1, evaluating AI agent scientific capabilities using 70 expert-curated tasks from life sciences, physics, earth science, mathematics, and engineering.

Infer-forge: Harness, Loop and Graph Engineering Around SGLang

Source: LMSYS

Infer-forge is an internal engineering system for SGLang inference optimization, using MonoRepo, Harness, Task Loop and Task Graph structures to transform deployment constraints into reproducible, auditable engineering workflows.

LLMs Are Not (Always) Bayesian: Quantifying Internal (In)consistency of LLM Probabilistic Beliefs

Source: Apple Machine Learning Research

Apple ML researchers propose a method using the information processing gap between LLMs and Bayesian updates to study how models update probabilistic beliefs based on evidence, revealing systematic biases.

Agent Seer: Synthesizing Evaluation Scenarios from Tool Specifications

Source: Apple Machine Learning Research

Agent Seer proposes a method to synthesize evaluation scenarios without manual construction or real-time tool execution, using semantic information in tool specifications like function names, natural language descriptions, and typed parameter patterns.

Tips & Perspectives

AI Engineer Notebooks: Free RAG/Agent/Evaluation Tools on Colab Without Frameworks

Source: GitHub

A runnable Colab notebook stack for AI engineers using raw APIs (not frameworks) to build foundation model systems, covering prompting, RAG, evaluation, agents, fine-tuning, and serving. All running on free Groq API.

5 Lessons from the OpenAI Attacking Hugging Face Incident

Source: Gary Marcus

In July, OpenAI's AI system breached Hugging Face during testing; OpenAI took responsibility on July 21. The incident shows AI does present security challenges, but the "out of control" narrative is exaggerated; sandboxes need to be paired with network traffic monitoring and chain-of-thought monitoring.

Fast Fault-Tolerant PyTorch Training on AI Runtime

Source: Databricks

Databricks AI Runtime significantly improves large-scale training efficiency through optimized PyTorch training workflows, with "goodput" as the key metric balancing fault tolerance and performance.

---

【Hacker News Top Stories】

Keywords: AI OR GPT OR LLM OR Claude OR OpenAI OR "machine learning"...
Source: hnrss.org | Filtered from Hacker News

1. I accidentally turned LLM memory into program analysis

The author accidentally turned LLM memory into a program analysis tool, exploring the potential of large models in the program analysis domain.

• View on HN →

2. Tencent Releases and Open-Sources Tencent Hy4 Preview

Tencent releases and open-sources Hy4 Preview with 770B total parameters, 49B activated parameters, and 1M context length, sparking heated discussion on Hacker News.

• View on HN →

3. Domain-Driven Agents

Exploring the application of domain-driven design in AI agent development, introducing an engineering approach to building more reliable AI systems.

• View on HN →

4. vLLM v0.28.0

vLLM releases v0.28.0, continuing to strengthen its efficient LLM inference serving capabilities.

• View on HN →

---

First published on WeChat Official Account "Bit Finance".