AI Morning Brief — 30 June 2026
DeepSeek V4 support merged into llama.cpp overnight, making the model runnable locally ahead of its mid-July official API launch; Meituan’s LongCat-2.0 simultaneously confirmed that large-scale pretraining on domestic Chinese hardware is viable at the trillion-parameter scale. Anthropic dominated the enterprise news cycle with its first US state government deal and a fresh security advisory for Claude Code users.
Top stories
- Anthropic signs California deal. Anthropic agreed to supply Claude to California state agencies at half the standard price under an agreement brokered with Governor Newsom — the company’s first major US state government contract. via TechCrunch – AI
- Amazon distilling Anthropic models. Amazon engineers are building smaller internal versions of Anthropic’s Claude before token-based pricing begins next year, with the company also exploring OpenAI as an alternative supplier. via THE DECODER
- DeepSeek V4 lands in llama.cpp. A community pull request merged overnight added DeepSeek V4 support (build b9840), enabling local inference; DeepSeek confirmed an official API launch in mid-July with peak-valley pricing. via llama.cpp releases
- LongCat-2.0 from Meituan. The Chinese food-delivery company released a 1.6 trillion-parameter MoE language model trained on roughly 50,000 domestic accelerators — a milestone for large-scale pretraining without Western hardware, previously the stealth owl-alpha model on OpenRouter. via Hacker News (front page)
- Cursor ships iOS app. Cursor launched a native iOS app in public beta, letting developers launch and monitor cloud coding agents from a phone; Composer 2.5 is offered at 75% discount through July 5. via Cursor
- Claude Code malware warning. Mozilla’s 0DIN security team demonstrated that Claude Code will execute runtime-fetched malicious code hidden in a GitHub repository’s setup flow — invisible to static scanners and to the AI agent itself. via THE DECODER
Who shipped
xAI shipped Grok voice APIs to the Vercel AI Gateway and launched hosted MCP servers on the X platform, giving AI tools a direct programmatic integration point for search, trends, and publishing. Anthropic brought Claude to NVIDIA GB300 Blackwell Ultra GPUs on Azure and rolled out Claude Opus 4.8 fast mode to GitHub Copilot as a preview. Meta published Brain2Qwerty v1 in Nature Neuroscience, demonstrating sentence decoding from non-invasive brain recordings; v2 model weights are publicly released. Google expanded Gemini’s personalized image generation to free users in the US.
Open-source pulse
DeepSeek V4 is now runnable locally after a community pull request landed in llama.cpp overnight. Ornith-1.0 from deepreinforce-ai arrived as an MIT-licensed self-scaffolding agentic coding model family (9B to 397B MoE, built on Gemma 4 and Qwen 3.5). National Design Studio released Rampart, a 14.7 MB in-browser PII redaction model that strips personal information client-side before data reaches any server.
Money, infra & hardware
Samsung and SK Hynix committed a combined $590 billion to expand memory chip production as AI data-center demand pushes HBM prices toward 50% quarterly gains through 2027; South Korea’s broader national plan targets $1 trillion in chips and humanoid-robot investment by 2028. Omen AI raised a $31 million Series A to monitor coolant chemistry and prevent bacterial outbreaks in GPU data centers. The model-evaluation startup Arena disclosed it is now a $100 million business, less than a year after launching its commercial evaluation service.
Quiet corners
Mistral and Cohere were absent from the overnight cycle. OpenAI‘s only visible activity was a teaser for Codex-related hardware due July 15; no model news appeared. Google DeepMind‘s direct output was limited to a blog explainer on full-stack AI strategy.
By the numbers
- 346 stories in 24 h across 40+ sources
- Most-mentioned model: Claude
- Most-mentioned lab: Anthropic
- Notable absences: Mistral, Cohere
Compiled by AI Feed’s editor from all 346 headlines published on 29–30 June 2026.