Model Releases & Benchmarks
Kimi K3 Open Frontier. Moonshot AI released Kimi K3, a 2.8T-parameter open-weight MoE model that outperforms Claude Opus 4.8 and nears GPT-5.6 and Fable 5 on benchmarks, featuring 1M-token context and multimodal input. Open weights are promised by July 27, with API pricing at $3/$15 per million tokens.
- Kimi's open model K3 nears GPT-5.6 Sol and Fable 5 while signaling the end of super cheap Chinese AI(opens in a new tab)
- Kimi K3, and what we can still learn from the pelican benchmark(opens in a new tab)
- [AINews] Kimi K3 2.8T-A50B: the largest open model ever released; Opus 4.8-class at Sonnet 5 pricing(opens in a new tab)
- Kimi K3 weights to be released on the 27th.(opens in a new tab)
Inkling Model Released. Thinking Machines Lab launched Inkling, an Apache 2.0-licensed 975B MoE model (41B active) trained on 45T multimodal tokens, aimed at fine-tuning rather than frontier performance.
Nemotron Embed Tops RTEB. NVIDIA open-sourced embedding models (8B and 1B) optimized for agentic retrieval; the 8B variant ranks #1 on the RTEB leaderboard, improving RAG accuracy.
Fugu Adds Nemotron. Sakana AI’s Fugu orchestrator now incorporates NVIDIA Nemotron models, showing that pooling multiple open models can collectively rival a single frontier system.
Local AI & Open Models
Giant Models Out of Reach. Kimi K3’s 2.8T size reignited debate: many hobbyists argue such models are effectively closed to home labs, while some plan elaborate server builds and others pause due to burnout. Optimists note that smaller 27B models may reach similar capability within months.
- Anyone else completely tuning out these massive "open weight" drops?(opens in a new tab)
- I’m taking a break(opens in a new tab)
- Kimi k3 is 2.8t! Will need to have an aggressive iQ2_XXS or IQ1.8!(opens in a new tab)
- How do you plan to run Kimi K3 locally?(opens in a new tab)
- Will we have a 27B model with Fable capabilities in 5 months? History says yes(opens in a new tab)
Speculative Decoding Stack. Combining DFlash with n-gram lookup drafting in llama.cpp speeds up Qwen 3.6 27B up to 6x on coding tasks with no quality degradation; MTP alone delivers 1.45x and DFlash 2.2x for chat or storytelling.
Budget Hardware Breakthrough. llama.cpp improvements pushed DeepSeek V4 Flash from 2 to 7 t/s on a single RTX 4060 Ti and CPU, demonstrating that recent optimizations are unlocking heavy models on modest gear.
Expert Prefetch Technique. A novel method uses the model’s MTP head to predict upcoming expert activations, potentially increasing CPU/GPU offload throughput from 30 to 150–200 t/s by hiding PCIe latency.
Developer Tools & Hardware
Firefox-in-WASM Demo. A team compiled Firefox to WebAssembly, enabling it to run inside Chrome via a WebSocket proxy; the AI-assisted build consumed an estimated $25K worth of tokens but cost far less under subscription plans.
Mermaid ASCII Tool. An existing Go library for converting Mermaid diagrams to ASCII art was ported to WebAssembly, now offering color output alongside an earlier Rust-based version.
OpenWiki Adds OKF. OpenWiki 0.2 introduces the OKF knowledge format, improving codebase documentation structure for coding agents and reducing the tokens needed for navigation.
DoorDash CLI for Agents. DoorDash released a command-line tool (dd-cli) that lets developers and AI agents search stores, find deals, and place orders programmatically, signaling the rise of agentic commerce.
Codex Micro Controller. OpenAI and Work Louder unveiled the $230 Codex Micro, a hardware controller with joystick and rotary dial that lets developers manage AI agents without typing, showing agent status via RGB lights.
Industry & Enterprise
Google Workspace AI Overhaul. Google renamed NotebookLM to Gemini Notebook with cloud code execution, updated Vids to let users star in AI-generated videos with Gemini Omni, and enabled AI Mode to directly interact with apps like Instacart, Canva, and YouTube Music.
- Create, edit and star in videos with two Google Vids updates(opens in a new tab)
- Google Vids now lets you star in your own AI videos(opens in a new tab)
- Google rebrands NotebookLM as Gemini Notebook and opens its search app to third-party integration(opens in a new tab)
- Google is renaming NotebookLM to Gemini Notebook(opens in a new tab)
- Google’s AI Mode now lets you link and interact with select apps(opens in a new tab)
- Connect more of your apps to Search(opens in a new tab)
Gemini Next-Gen Delayed. Google’s upcoming Gemini model has been pushed back after failing to meet internal quality goals, according to a Reddit report.
Roblox AI Game Builder. Roblox launched Build, an AI feature that creates playable games from text prompts on mobile, handling everything from environment to sound using a mix of open and proprietary models.
AI-Driven Energy IPOs. Energy firms raised $12.6B in H1 2026—the most since the dot-com era—as investors bet on the surging electricity demands of AI data centers.
Linus Backs AI in Kernel. Linus Torvalds declared Linux will not be anti-AI, telling those opposed to tools like the Sashiko agentic code reviewer to fork the kernel if they dislike AI-assisted contributions.
Workers Strike Over Robots. Hyundai auto workers in South Korea staged the first strike over humanoid robot deployment, protesting plans for 25,000 Boston Dynamics Atlas robots in Ulsan.
Visual AI Startup Raises $55M. Elorian, founded by ex-DeepMind researcher Andrew Dai, secured a $55M seed round at a $300M valuation pre-product, targeting visual reasoning for AI systems.
World Model Startup Eschews Hype. AMI Labs CEO Alexandre LeBrun rejects terms like AGI and superintelligence, focusing instead on practical world-model applications for robotics and manufacturing.
OpenAI’s Hardware and Merch. OpenAI launched a $230 Codex Micro controller for agentic workflows and a $70 ChatGPT basketball, combining developer tools with consumer marketing.
Compute Spending Outpaces Visibility. VentureBeat reports enterprises are buying AI compute faster than they can measure unit economics, with GPU utilization often below 50% and many planning provider switches.
Agent Trust & Readiness Gaps. Surveys find 54% of enterprises have had agent security incidents, yet most share credentials; context inaccuracies lead to confident wrong answers; and evaluation methods don’t reflect real-world outcomes, though production rollouts continue.
- The agent security gap: 54% of enterprises have already had an AI agent incident, and most still let agents share credentials(opens in a new tab)
- The AI context gap: Enterprise AI organizations have a trust problem, not a retrieval problem — and most are still building the fix(opens in a new tab)
- The agent evaluation gap: Enterprise AI organizations have a reality-alignment problem, not a coverage problem — and most are shipping to production anyway(opens in a new tab)
Policy & Society
EU Opens Android AI. The EU ordered Google to allow competing AI platforms on Android and share search data, leveraging DMA rules to break Gemini’s preferential access.
xAI Sues CSAM User. After refusing to help police, xAI sued a user who generated child sexual abuse material with Grok, following a class action and public pressure over its inability to prevent such outputs.
Germany Regulates AI Search. German media regulators classified AI chatbots as content providers, ruling that Google’s AI Overviews and Perplexity can be held liable for generated outputs under state media law.
NY Governor AI Law Review. New York Governor Hochul used AI to analyze all state regulations, completing a review in months that would have taken staff years, while also imposing a moratorium on new hyperscale data centers.
OpenAI’s Teen Safety Pitch. OpenAI argued teens deserve access to AI with tailored protections, highlighting ChatGPT’s role in education while noting safeguards for younger users.
Open Weights Alarms Continue. Amid Kimi K3’s release, Anthropic’s Dario Amodei is expected to renew warnings about the dangers of open-weight frontier models.
That's everything for this day – about a 5-minute read.