- DeepSeek V4 goes stable: DeepSeek shipped V4-Pro (1.6T total / 49B active) and V4-Flash (284B / 13B active) with a 1M-token context now default across all official services, DeepSeek Sparse Attention, and OpenAI- and Anthropic-compatible APIs — the legacy
deepseek-chatanddeepseek-reasonermodels retire after today. - OpenAI launches Presence + Health in ChatGPT: Presence (July 22) is a governed enterprise agent platform connecting agents to internal systems with policies and guardrails (early partners BBVA, SoftBank, IAG), and Health in ChatGPT (July 23) pushes ChatGPT into a high-trust consumer category.
- Washington enters the Kimi K3 fight: OSTP Director Michael Kratsios alleged Moonshot distilled Anthropic's model to build the ~2.8T-parameter Kimi K3 and routed restricted Nvidia GB300 compute through Thailand; free weights are still due July 27.
- Google refreshes Gemini Flash: Gemini 3.6 Flash (the new workhorse, ~17% fewer output tokens), 3.5 Flash-Lite, and a security-focused 3.5 Flash Cyber shipped July 21 — but the anticipated Gemini 3.5 Pro refresh did not.
AI News (Top Updates)
1. DeepSeek V4 lands as a stable open-weight release
DeepSeek pushed V4 to a stable release, split into V4-Pro (1.6-trillion-parameter MoE with 49B activated) and the leaner V4-Flash (284B total / 13B active). A 1-million-token context is now the default across every official DeepSeek service, and the models use Token-wise compression plus DeepSeek Sparse Attention (DSA) for efficiency. Both speak the OpenAI ChatCompletions and Anthropic APIs and are tuned for agentic coding (Claude Code, OpenCode, OpenClaw); the legacy deepseek-chat and deepseek-reasoner endpoints retire after July 24.
2. OpenAI launches Presence, a governed enterprise agent platform
Announced July 22, OpenAI Presence connects AI agents to a company's internal systems with governance, policies, permissions, and guardrails, aimed at customer support, sales, and higher-risk workflows. It shipped as a limited general-availability product, with BBVA, SoftBank, and IAG named as early partners exploring deployments. The move puts OpenAI squarely into the integration-and-deployment layer, not just the model layer.
3. OpenAI launches Health in ChatGPT
On July 23, OpenAI introduced a dedicated Health experience inside ChatGPT, published under its Product category. It lands the same week as a ChatGPT for Small Business program (July 21), two new board appointments (David Vélez and Robin Vince), and a jointly disclosed security incident with Hugging Face — a busy stretch of consumer, enterprise, and safety news out of the OpenAI newsroom.
4. The White House enters the Kimi K3 distillation dispute
OSTP Director Michael Kratsios publicly alleged that Moonshot AI distilled Anthropic's model to build Kimi K3 — a ~2.8-trillion-parameter model that reportedly topped the Frontend Code Arena with a 76% win rate — and accessed restricted Nvidia GB300 servers routed through Thailand. Independent analysis (Ryan Greenblatt, Redwood Research) noted K3 ‘identifies as Claude disproportionately often,’ while flagging innocent explanations like web-data contamination or prompt leakage. Moonshot has not conceded, and free weights are still scheduled for July 27.
5. Google refreshes Gemini's Flash tier — but skips 3.5 Pro
On July 21, Google released three new models: Gemini 3.6 Flash (its new default workhorse, with improved coding, knowledge, and multimodal performance using up to 17% fewer output tokens), Gemini 3.5 Flash-Lite (the cheapest general-purpose tier), and Gemini 3.5 Flash Cyber (a security-vulnerability specialist limited to governments and trusted partners). The anticipated Gemini 3.5 Pro refresh — last updated in February — did not ship, reportedly because internal targets were not yet met.
6. Alibaba previews Qwen3.8-Max, a 2.4T-parameter flagship
Alibaba unveiled a preview of Qwen3.8-Max (July 19), a 2.4-trillion-parameter sparse-MoE multimodal model that processes text, images, video, and documents. Alibaba claims it should outperform Qwen3.7-Max on coding, full-stack development, and office workflows and positions it as ‘second only to Fable 5’ — though there's no public benchmark table, model card, or license yet. Preview access runs through the Token Plan at roughly 10% of standard pricing, with open weights promised ‘soon.’
7. Microsoft and Mistral expand a sovereign-AI partnership
On July 21, Microsoft committed a multibillion-dollar European compute buildout (thousands of Nvidia Vera Rubin GPUs for Mistral) and brought Mistral's Medium 3.5 and OCR 4 models into Microsoft Foundry, with Medium 3.5 also in Copilot Studio. Azure lets organizations run the models cloud-scale, cloud-connected, or fully disconnected, keeping control over data and operations for regulated industries.
8. Meta ships Muse Image, its first Superintelligence Labs image model
Meta introduced Muse Image, the first image-generation model from Meta Superintelligence Labs. It generates and edits images from conversational prompts, blends multiple photos into one, renders legible in-image text for guides and infographics, and powers 30+ effects in Instagram Stories plus room-redesign features with real retail products. It's live now in Meta AI, with Instagram Stories and WhatsApp available in limited countries and Facebook, Messenger, and Advantage+ for advertisers coming soon.
9. Nvidia's Vera Rubin ramps as Japan builds a national AI factory
Nvidia said Vera Rubin production is ramping at CoreWeave, Google Cloud, Microsoft Azure, and Oracle Cloud (July 21), pairing it with the new Spectrum-6 networking for gigascale AI factories and citing best-in-class performance-per-watt and lowest token cost. Days earlier (July 16), Japan's government and industrial leaders launched what Nvidia calls the world's first national AI infrastructure — a Vera Rubin factory with 13,750 Vera CPUs and 27,500 Rubin GPUs — while new Jetson Thor computers bring foundation-model inference to the robotics edge.
10. xAI ships Grok 4.5, open-sources Grok Build, and adds Automations
On July 16, xAI released Grok 4.5, its most capable model ‘built for coding, agentic tasks, and knowledge work,’ alongside a new Automations feature for scheduled autonomous tasks. A day earlier it open-sourced Grok Build, its coding-agent and terminal tool, and it had already added 21 new flagship voices on July 6. Elon Musk has since teased a roughly 2-trillion-parameter Grok 4.6 in training.
11. Anthropic: Claude for Teachers, an Economic Index connector, and $20M to Public First
Anthropic launched Claude for Teachers (July 14), giving verified U.S. educators premium Claude, alongside a $10M commitment to Canadian AI research. On July 22 it shipped a no-install connector letting anyone query the Anthropic Economic Index conversationally, paired with a new Economic Futures Research Fund agenda. It also donated another $20M to Public First Action (July 21) and opened rare-disease research grants under its AI for Science program — all building on the June 30 launch of Claude Sonnet 5.
12. Microsoft 365 Copilot brings MCP agents into Office
Microsoft's mid-July Copilot wave (July 15) made Model Context Protocol agents usable inside Word, Excel, PowerPoint, Outlook, and Catalyst, and let admins publish Agent Builder agents to the Agent Store under review. The release also added tenant-wide Copilot Prompt Galleries, auto-generated brand kits from uploaded guidelines, and hierarchical ACLs for Confluence and ServiceNow connectors.
13. Hugging Face ecosystem: LeRobot, ML Intern, and an open-weight surge
Nvidia and Hugging Face expanded the open-robotics LeRobot project with new models and frameworks, while DeepSeek V4 (stable today) and Kimi K3 (weights July 27) converge on the Hub as two of the year's largest open releases. Hugging Face's own ML Intern — an open-source agent that searches papers, picks datasets, writes code, and runs training — reportedly scored 32% on GPQA versus Claude Code's 22.99%, and installs via a simple CLI.
14. Business & policy: OpenAI's $30B Georgia campus and a UN governance push
OpenAI detailed ‘Project Camellia,’ a ~$30B, 1,400-acre, ~3.2-gigawatt data-center campus in Effingham County, Georgia, with power phased in 2028–2032, $80M in community benefits, and $71M in Codex credits for Georgia students. Meanwhile the UN issued a fresh push for AI governance amid warnings of potential ‘catastrophic harm,’ and export controls face a credibility test as third-country compute rental emerges as a workaround. The White House is expected to unveil a frontier-AI framework before August 1.
Top 5 New / Popular AI Products
1. DeepSeek V4-Pro
Open-Weight Frontier ModelDeepSeek's stable 1.6-trillion-parameter MoE (49B activated) with a 1M-token default context and DeepSeek Sparse Attention, speaking both OpenAI and Anthropic APIs and tuned for agentic coding.
2. Gemini 3.6 Flash
New Flash ModelGoogle's fast, cheap frontier-class workhorse (July 21) with improved coding, knowledge, and multimodal performance using up to 17% fewer output tokens than 3.5 Flash.
3. OpenAI Presence
Enterprise Agent PlatformA governed platform (July 22) that wires AI agents into internal systems with policies, permissions, and guardrails, in limited GA with BBVA, SoftBank, and IAG as early partners.
4. Qwen3.8-Max
Preview Flagship ModelAlibaba's 2.4-trillion-parameter multimodal MoE preview (July 19), claimed ‘second only to Fable 5,’ available at ~10% of standard pricing via the Token Plan.
5. Meta Muse Image
New Image ModelMeta Superintelligence Labs' first image model — conversational generation and editing, multi-photo blending, and legible in-image text — live now in Meta AI, Instagram Stories, and WhatsApp.
Discussion