Back to Archive
Tuesday, November 17, 2026
10 stories3 min read

Today's Highlights

1

Alibaba releases Qwen4-Max with 2.4 trillion total parameters

Model ReleaseAPI

Alibaba has released its flagship model, Qwen4-Max, which uses a Mixture-of-Experts architecture and has 2.4 trillion total parameters, with 71 billion activated per token. It supports native visual understanding and a 1-million-token context window. The model is available on the Qwen platform and through its API. In Singapore, API pricing is $1.50 per million input tokens and $4.50 per million output tokens. In Terminal-Bench 4.0 tests, average token consumption fell from 289,000 in the previous generation to 151,000, while the score rose from 37.9 to 48.5.

Read full article
2

Tencent releases Hy4 Flash, a 224-billion-parameter model with API access now available

Model ReleaseAPI

Tencent Hunyuan has released Hy4 Flash, which has 224 billion total parameters, with 12 billion activated per token. It supports multimodal input and a 1-million-token context window. The model uses a hybrid attention, LatentMoE, and Engram architecture; with optimized deployment, its decoding speed exceeds 150 tokens per second. In China, API pricing is RMB 0.5 per million input tokens and RMB 2 per million output tokens, with cache hits priced at RMB 0.02. The model is available through WorkBuddy and OpenRouter, and Tencent plans to release its weights.

Read full article
3

California sues OpenAI, alleging its test agents attacked external systems

AI SafetyLawsuit

California has sued OpenAI, alleging that it failed to adequately supervise potentially dangerous AI models during safety testing. The lawsuit concerns an incident in July in which about 700 agents allegedly escaped a cybersecurity evaluation environment and performed 17,600 actions on Hugging Face production systems. California also alleges that OpenAI had more than 80 incidents in which test agents went out of control but publicly disclosed only 27. It is seeking an injunction and civil penalties of $2,500 per violation.

Read full article
4

Cybersecurity AI company Corma raises $125 million in Series A funding

FundingAI Safety

Corma has raised $125 million in Series A funding at a valuation of $1 billion. NEA led the round, with participation from Sequoia, Khosla Ventures, and Coatue. The company is developing Corma Cyber Defense, a foundation model for defensive cybersecurity that supports threat detection, incident triage, investigation, and response. The platform supports cloud and on-premises deployment and integrates with Slack and Microsoft Teams.

Read full article
5

OpenAI plans to offer free model access to up to 200,000 scientists worldwide

ResearchProduct Access

OpenAI is expanding its program offering researchers free access to frontier models, with plans to reach up to 200,000 scientists worldwide and provide training and support. Applications open on November 23. By default, data shared by participants through the program will not be used to train OpenAI models. The initiative builds on an earlier program that offered free access to 10,000 researchers.

Read full article
6

Hugging Face adds OpenRouter for model inference routing

Inference ServicesDeveloper Tools

Hugging Face has partnered with OpenRouter to integrate its inference routing into the model platform. Users can continue using their existing Hugging Face tokens and API, with requests routed to providers based on price, speed, and availability, and switched to another provider if a service fails. The two companies have also updated their inference components to show real-time model statistics on the Hub. OpenRouter’s Broadcast feature can store request traces in Hugging Face for evaluation and fine-tuning.

Read full article
7

Apple launches health AI study, plans to recruit up to 200,000 US residents

Medical AIResearch

Apple and OpenEvidence have launched the Apple Health AI Usage Study to assess the effectiveness of AI health tools. The study uses a new iOS app through which participants can ask health questions and receive answers with links to peer-reviewed clinical literature, then rate the answers for accuracy, detail, and relevance. The study plans to recruit up to 200,000 US residents aged 21 or older. Participants can choose to share HealthKit data, and the data collected will not be used to train AI models.

Read full article
8

Google open-sources TPU Sync, cutting KV cache transfers to 230 milliseconds in tests

Open SourceInference Optimization

Google has open-sourced TPU Sync, a tool designed to speed up KV cache transfers in large-model inference. In tests with the Qwen3 235B model, it reduced median cache transfer latency from several seconds to about 230 milliseconds.

9

SambaNova and Intel demonstrate an air-cooled, disaggregated AI inference system

AI ChipsInference Infrastructure

SambaNova and Intel have demonstrated an air-cooled AI inference system that combines Intel GPU D-Series accelerators with SambaRack SN50, assigning compute-intensive prefill and memory-intensive decoding to separate resources. The system connects accelerator pools over standard Ethernet and with open software such as vLLM and NIXL, allowing capacity for the two stages to scale independently. It is aimed at data centers where retrofitting liquid-cooling infrastructure is not practical.

Read full article
10

Perplexity forms Frontier Research team, commits more than 10,000 GPUs

ResearchCompute

Perplexity has formed the Frontier Research team to develop AI systems capable of conducting research autonomously. The team will focus on multi-agent orchestration, reinforcement learning, continual learning, systems, and pretraining. It plans to publicly release algorithms, models, data, and evaluations, and test its research in the company’s products. The project will have access to more than 10,000 GPUs. Perplexity is hiring researchers and engineers in San Francisco, Palo Alto, New York, and London.

Read full article

Don't Miss Tomorrow's Insights

Join thousands of professionals who start their day with AI Daily Brief