China's domestic chips run 2.8T Kimi K3 on day zero
Alibaba Cloud says its Zhenwu M890 supernode is the first Chinese system to run Moonshot's 2.8-trillion-parameter model; Huawei Ascend also claims day-0 support.
July 27, 2026
All briefings and quick hits published that day.
- ⚡ Quick hits
- AIGC Meshy demonstrated clean cel-shaded 3D rendering in Blender driven by its 8K generated textures. X @MeshyAI ↗
- ⚡ Quick hits
- AIGC Recraft shipped V4.1, showcasing jewelry renders with physically convincing glass refraction and depth across rings, earrings and collars. X @recraftai ↗
Samsung engineers defect to SK Hynix over HBM bonuses
A reported 81.5% of Samsung foundry staff want out within two years as rival SK Hynix pays bonuses more than three times larger on the back of AI memory profits.
Report: most top Hugging Face editors will undress women
European nonprofit AI Forensics found seven of the nine most popular image-editing models on Hugging Face complied with simple undressing prompts.
- ⚡ Quick hits
- Research Shanghai AI Lab's InternLM introduced Visual Pretraining, a scalable paradigm letting foundation models learn directly from rendered visual documents. X @intern_lm ↗
Cursor launches ₹649 India tier with cloud agents
Cursor's first localised plan bundles Composer 2.5, Grok 4.5, autonomous cloud agents and an iOS app for roughly a third of Pro's price.
- ⚡ Quick hits
- Research Apple detailed the audio detokenizer behind Siri Expressive Voices: decoupled temporal-depth diffusion streaming at ~21MB peak memory, 16x faster than real time, +0.28 MOS. Apple ML Research ↗
- Research An engineering writeup claims a $500 reinforcement-learning fine-tune of a 9B open model beat frontier models on a product-catalog review task. Fermisense ↗
- Funding Coursera invested $100 million for a one-third stake in LearnVector, Andrew Ng's new AI-native tutoring company, with first products targeted for early 2027. Business Wire ↗
- AIGC KAIST's ReDesign recovers editable design files from raster images via agentic tool use, benchmarked on 909 Figma files and nearly 15,000 edit instructions. arXiv ↗
Anthropic says it never sought a ban on open-weights models
Responding to months of speculation, Anthropic published its open-weights position: no bans, but mandatory pre-release safety testing for any sufficiently capable model.
- ⚡ Quick hits
- Industry Simon Willison published an opinionated guide mapping which AI tool to reach for across common everyday tasks. Simon Willison ↗
- Industry Satya Nadella warned that companies betting everything on a single AI provider may not survive, urging multi-model strategies. TechCrunch ↗
- Agents OpenAI's GPT-5.6 Terra and Luna xHigh variants entered Agent Arena at #15 and #17, well behind sibling Sol, which ranks top two. X @arena ↗
- Agents Cursor added Kimi K3, saying it scores close to the frontier on CursorBench and runs on US-based inference from Fireworks, Together, and Baseten. X @cursor_ai ↗
Moonshot open-sources AgentENV, the sandbox fleet behind K3
The Kimi team and kvcache-ai released the Firecracker-based environment system used for Kimi K3's agentic RL training under an MIT licence.
- ⚡ Quick hits
- Safety Users found publicly shared Claude chats and Artifacts indexed by Google, exposing conversations people assumed were only link-accessible. TechCrunch ↗
- Industry OpenAI published research on small businesses finding staff increasingly use ChatGPT to handle work outside their own job descriptions. X @OpenAI ↗
- Models LMArena added a factuality ranking blending human preference with factual accuracy, where Claude Opus 5 with Max reasoning takes the #1 Text Arena spot. X @arena ↗
- Industry Verizon signed a roughly $1 billion dark fiber deal to connect Google data centers, calling it the first of many AI-driven contracts. Ars Technica ↗
Moonshot opens PerceptionBench; no model clears 60%
The Kimi team released a benchmark that strips reasoning out of multimodal evaluation and grades whether a model actually saw the image correctly.
- ⚡ Quick hits
- Policy A judge rejected Google's attempt to use DMCA claims against a web scraper, with the winning company declaring that Google and Reddit "do not own the Internet." Techdirt ↗
- Open Source Together AI became a day-zero launch partner for Kimi K3, offering immediate high-throughput developer access to the model. X @Kimi_Moonshot ↗
- Models OpenAI rolled GPT-Live in ChatGPT Voice out globally to Edu, Business, and Enterprise plans. X @OpenAI ↗
- Policy ChatGPT now refuses direct requests to write in a named living author's style, though testers say it still reproduces a similar feel indirectly. Ars Technica ↗
Microsoft's MAI-Cyber-1-Flash halves cost of its MDASH harness
Microsoft's first in-house security model handles up to 90% of vulnerability-hunting work inside MDASH, with OpenAI models still reserved for the hardest cases.
- ⚡ Quick hits
- Industry Meta opened Threads direct messages to Meta AI, letting users chat with the assistant inside their DM inbox. TechCrunch ↗
- Agents LlamaIndex released create-llama-worker, a scaffold that deploys a LlamaParse document-parsing worker to Cloudflare Workers in minutes. X @llama_index ↗
- Agents Perplexity shipped pplx, a single-binary CLI exposing only web search and page fetch as strict JSON, built for coding agents rather than chat. MarkTechPost ↗
- Industry New data shows Google's AI-generated search results are rapidly becoming the default experience for most queries. TechCrunch ↗
- Agents Cognition added Kimi K3 to Devin Desktop and CLI, saying it scores 58.2% on FrontierCode 1.1 Extended with a 63.6% pass rate. X @cognition ↗
- AIGC PixVerse upgraded the speech engine inside its Avatar Lip Sync Mini App for more natural talking-head video. X @PixVerse_ ↗
- AIGC PixVerse added Depth Map control to its Mini Apps, letting creators drive image-to-video motion with a grayscale depth video extracted from a reference clip. X @PixVerse_ ↗
- Agents Kling published a tutorial for producing a full beauty-clinic ad end to end through its MCP server, from concept to finished cut. X @Kling_ai ↗
METR proposes 'expenditure horizon' to price agents vs humans
METR's new metric finds the budget at which AI research agents stop being cheaper than people — currently in the low four figures for only the newest models.
Nvidia-led Open Secure AI Alliance launches with 50+ members
Nvidia's Open Secure AI Alliance lists Microsoft, IBM, Hugging Face and others among 50-plus inaugural partners; separately, the open-weights policy letter has grown to 50 signatories including Google and OpenAI, with Anthropic and Amazon still out.
Moonshot's 2.8T-parameter Kimi K3 weights land on Hugging Face
Moonshot AI has published Kimi K3's full 2.8-trillion-parameter weights under the Kimi K3 License — the largest open-weight model release to date.