- Reuters names CoreWeave in industry warning over proposed Chinese optics ban
Reuters reports CoreWeave joined Amazon, Google, Meta, Cisco and Microsoft in warning compliance could take up to a decade, while officials were recently said to be moving forward; Reuters gave no date for that communication, and the scope and timetable remain unconfirmed.
- vLLM quantization cuts Qwen3-8B memory by 61% in contributor test
Merged on November 16th, vLLM's load-time weight-only quantization boosted single-request speculative decode throughput by 13.1% and 14.1% on two bandwidth-bound GPUs in the contributor's tests.
- Verisource finds agent endpoints on eight of 10 sampled Shopify stores
Verisource's November 16th probe found active UCP endpoints on eight sampled Shopify stores and no custom '/agents.md' at the 10 tested domains, with five serving stock Shopify files.
- Tencent launches Hy4 flash with input tokens listed at $0.070 per million
Tencent's 224B-parameter model activates 12B parameters per token, supports 1M tokens of context, and is priced well below Hy4 for routine WorkBuddy tasks.
- Microsoft merges one-agent-per-container change for .NET Foundry hosting
Apps adopting the merged code will need separate containers for externally served agents; specialist handoffs can continue behind one workflow agent.
- Anthropic can buy about 875,000 leased TPUs for $95 after $36B in rent, FT reports
The Financial Times says Anthropic's $7.2B-a-year term covers the first 1 GW batch for four sites, inside $900.9B of commitments disclosed before its IPO.
- llama.cpp patch reports 25% Qwen3.5 prefill gain in MI100 test
The merged change raised Q5_K_M throughput from 422.29 to 529.34 tokens per second at a 512-token microbatch in the contributor's single-GPU test.
- YTL Power falls nearly 10% as US chip rules roil regional shares
Reuters reported YTL Power down nearly 10% on November 17th as investors assessed licence requirements for covered chips sent to Singapore, Malaysia, Thailand, Indonesia and Brazil.
- vLLM test reports 13% higher Step5 throughput with PrimTS on Blackwell
The merged, opt-in FlashInfer path reported 147.04 tokens per second at roughly 96K input and concurrency 32, and 5.438-second time to first token for a separate 160K batch-one case on four GB300s.
- ONNX Runtime cuts single-layer DeepSeek-style CPU decode to 3.45 ms in test
The dedicated operator also halves the tested FP16 cache to 1,152 bytes per token per layer.
- You.com lists Alibaba's Qwen4-Max from the model's launch day
The November 16th flagship is live through You.com's Inference Router at $1.35 per million input tokens, with web-search grounding available through You.com Search.
- OpenAI says scientists estimate ChatGPT nearly doubles research speed
OpenAI based the 1.9-times estimate on 251 surveyed users in its free academic program, then expanded that program toward 200,000 researchers.
- AWS defaults unpinned AgentCore harnesses to Claude Opus 5 in commercial Regions
AWS lists the change under November 2026 in its release notes without giving a day. In supported commercial AWS Regions, an API-created AgentCore harness without explicit settings uses Claude Opus 5, task instructions, context summarization and a managed built-in tool set that AWS may revise.
- LiteLLM opens its multi-gateway control plane to scripts and automation
Dated November 16th, 2026, BerriAI's announcement says the Enterprise-licensed API brings multi-deployment management previously handled in a UI to scripts and automation, with dataplane registration, fleet usage queries and management proxying.
- Corvex expects to fund $14M of $103M Blackwell deal's deployment
The three-year deal would lift Mid-Atlantic 1 contracted revenue to about $56M a year once more than 1,000 B200 systems go live in early 2027.
- Howie joins Google Labs to build AI scheduling tools for small businesses
The Seattle startup founded by Austin Petersmith and Dave Newman says existing customers will keep being served, two weeks after Axios reported late-stage Alphabet acquisition talks.
- SCOOP: OpenAI adds a remote switch for how Codex models run agent teams
The merged change lets a remote model catalog expose or withhold spawn, messaging, waiting and interruption tools for each Codex model.
- Tradeweb's TARA AI assistant enters government bond trading tests
About a dozen global government bond firms are testing TARA, five months after its U.S. credit debut, as Tradeweb aims for a broader release in the coming weeks.
- NVIDIA fixes B300 race that sank DeepSeek-R1 test accuracy to 0.53%
In a contributor's B300 test, requests completed despite garbage output from DeepSeek-R1 after a warp read stale NVFP4 scale factors.
- CyberNative finds silent model fallback in VS Code subagents as Copilot retires models
CyberNative counted 360 public agent files pinned only to models scheduled for retirement on November 16th, with an annual-plan exception. Its VS Code review found subagents could inherit the main chat model without an in-chat notice.
- xAI workshop documents Cursor /goal agents that keep starting new turns
A November 16th runbook in xAI's public workshop repo spells out the cloud lifecycle: work past a closed laptop, wait for CI, and archive the agent to end it.
- vLLM shrinks a 970 MB output head to 300 MB and gains 12% on Strix Halo
Xiao Yu's merged patch compresses Qwen3.6's 16-bit lm_head at load, cuts median decode time by 2.82 ms, and publishes the perplexity cost beside the gain.
- Hugging Face hands its open-model inference routing to OpenRouter
Clement Delangue's platform now routes chat model calls through Alex Atallah's OpenRouter, adding failover, provider controls and trace storage on the Hub.
- Querit lists its AI search API in AWS Marketplace to reach enterprise buyers
The Singapore search company, founded in 2025, can now sell its agent retrieval API through existing AWS accounts and procurement workflows.