Hey, it's Andreas.
Quick one before we start.

The Agentic AI Cohort 2 is open - and already a third full!

We start in September. Cohort 1's still running and people are building real things (one project even landed in the App Store🤯). Early-bird is capped at the first 20 seats → all info here and if you have questions send me a DM.

And one for tomorrow: DataCamp is hosting a free live panel on our new book, The Generative AI Career Masterplan — Tue, July 21, 11AM ET / 5PM CET.

Now, on to this week.

In today's issue:

Let's go

🌀 Thinking Machines Lab releases Inkling, its first open-weights model → 975B parameters (41B active), multimodal, 1M-token context, Apache 2.0 — the largest US open-weights model so far. The company says outright it isn't the strongest model available.

🌀 Bloomberg reports Gemini 3.5 Pro is months behind schedule → A late-June training-data update returned disappointing coding results, and Google's last Pro-tier model shipped back in February. Also cited: duplicated coding-tool efforts across Cloud, DeepMind and Android, and engineers who believe important code should stay human-written.

🌀 OpenAI ships Codex Micro, a $230 keypad for steering coding agents → Built with Work Louder: six RGB agent keys that track live status, a dial for reasoning level, 32 swappable keycaps. OpenAI's first hardware product, and a bet that running several agents at once has become a physical-workspace problem.

🌀 OpenAI ships ChatGPT Work, an agent that stays on a project for hours → Hand it a goal, approve the big moves, get back finished docs, sheets, slides and web apps. Same pitch as Anthropic's Cowork, but six months later.

🌀 Google renames NotebookLM to Gemini Notebook and adds code execution → Every notebook gets a secure cloud computer, so it can write and run code against your own sources instead of only summarizing them.

🌀 Anthropic ends the Fable 5 access saga → Fable stays in Max and Team Premium at half the usual limits. Pro and Team Standard get a one-time $100 credit, then pay-per-use at $10 in / $50 out per million tokens.

🌀 Mirai Labs open-sources Weaver, a faster speculative decoder for local models → Local inference runs at batch size 1, so drafts have to get large without acceptance collapsing. Weaver's hybrid factorized-plus-autoregressive drafter reports 4.37× over autoregressive decoding and +24.7% on the strongest public baseline. Paper, SGLang kernels and weights all shipped together.

📘 Google PM open-sources Awesome LLM Apps, 100+ runnable agent and RAG projects → Apache-2.0, 125K stars, sorted by pattern: starter and advanced agents, multi-agent teams, voice, memory, MCP, plus 21 RAG tutorials.

📘 DeepLearning.AI teaches fast LLM inference on Cerebras hardware → Free, 1h 42m, intermediate: ten lessons on why wafer-scale beats GPUs on latency, plus real-time personalization, multi-tool workflows and agentic coding.

📘 OpenAI publishes prompting guidance for GPT-5.6 → The headline finding: stripping repeated instructions and simplifying tool descriptions raised eval scores 10-15% while cutting tokens 41-66% in OpenAI's own tests.

📘 Rohit Ghumare open-sources a 503-lesson AI engineering curriculum → Twenty phases, MIT licensed, 40.2K stars, four languages: Python, TypeScript, Rust, Julia.

📘 Anthropic gives US K-12 teachers Claude free for a year → Verified American K-12 educators get premium Claude including Claude Code and Cowork, a teaching skills library, curricula mapped to all 50 states' standards, and nine integrated edtech tools. Sign-ups run until June 2027.

P.S. Got a good new course? Send it my way on LinkedIn or just hit reply.

🔹 Anthropic opens talks to rent Meta's spare AI compute → Reportedly worth up to $10 billion over two years, proposed by Anthropic in June and still early enough that it may not close. If it lands, Meta becomes landlord to a direct rival.

🔹 Boris Cherny maps five levels of AI-native engineering adoption → From gated (zero agents) to AI-native (1,000+, most of them kicked off by Claude itself), with Anthropic sitting at level three org-wide.

🔹 Hassabis proposes a FINRA-style standards body for frontier AI → Labs would hand over models 30 days before release for testing on cyber, bio and deception risks — voluntary at first, mandatory once the protocol proves out. He wants it running before year-end.

🔹 Thinking Machines argues the future needs many AIs, not one → A Hayekian case: useful knowledge is tacit, local and fleeting, so a model trained in a handful of places and then frozen loses most of it.

🔹 Weco reports early evidence of recursive self-improvement → An outer agent rewrote the inner research agent's code and search strategy across 100 steps over eight unsupervised days, yielding seven successively better versions that beat a two-year hand-tuned baseline on three benchmarks.

🔹 OpenAI's CFO proposes "useful intelligence per dollar" over per-token pricing → Sarah Friar's scorecard asks four things: is the work useful, what does each successful task cost, can people depend on it, does value grow with scale.

🔹 Amplify's 2026 AI Engineering Survey maps how 1,053 teams build → Agent usage roughly doubled year over year, and 90% of agent-using teams now let agents write data, up from 52%. Cost is the new ceiling: 40% say it regularly limits how ambitiously they use AI.

Until this week, the frontier belonged to Claude and OpenAI's models. Open-weight alternatives were cheaper, sure, but never quite there, and the gap always showed up where it hurts: the last 20% of a real task. There are clear signs that Kimi K3, from Moonshot AI, changed that.

I have been trying to figure out why it feels like a bigger deal than its benchmark scores suggest, and I think I finally have a frame for it: the distance between the best model in the world and the best model you can download has never been this small. Call it the download gap. For two years it was measured in generations. As of this week it is measured in weeks.

A lot of quick conclusions are being drawn about K3 from saturated benchmarks and Elo charts rather than from testing it on genuinely hard problems. I did test it, and a few things stood out.

I tested it two ways: heavy coding work, long agentic sessions navigating a large repository and orchestrating terminal commands without losing the thread, and long-context work, dropping a full document set on it and asking questions that require connecting details hundreds of pages apart. Especially on the coding, it held state across sessions where most models start to drift. It matches many American frontier models and sometimes beats them, though it does not feel quite at Fable 5 or GPT-5.6 Sol level. Which is roughly where I expected a strong Chinese lab to be by mid-2026: an arrival on schedule, not a DeepSeek R1 surprise. It will be treated as one anyway, for reasons that have more to do with markets than capability.

The public evidence points the same direction. Within a day of launch, K3 sat at number one on Arena's Frontend Code leaderboard with 1,679 points, ahead of Claude Fable 5, in blind tests where developers picked the output they preferred. That is a 17-place jump from Kimi's last model.

Under the hood it is a 2.8 trillion parameter Mixture-of-Experts model (Moonshot rounds up to "3-trillion class," which is forgivable marketing), with 896 experts of which only 16, roughly 50 billion parameters, fire per token. The weights land on July 27. Nothing this large has ever been released openly.

On overall capability it is more measured: third on Artificial Analysis's Intelligence Index, roughly level with Claude Opus 4.8, behind only Fable 5 and GPT-5.6 Sol. Moonshot itself says the same, which I find more credible than the usual launch-day chest-beating.

One caveat matters more than the rankings here. Artificial Analysis found K3's hallucination rate regressed even as accuracy improved, so it gets more answers right while also confidently inventing more wrong ones, which happens to be the failure mode that shows up most in real work and least on leaderboards.

The thing early users keep pointing to is not a score. K3 is natively multimodal, and Moonshot wired that into what it calls vision in the loop: the model writes code, renders a screenshot of its own output, looks at it, notices the misaligned button or the muddy color scheme, and ships another version. It can see what it makes. As far as I can tell, that closed loop is what separates plausible code from finished software, and it shows most clearly in front-end work, the actual look and feel of apps and websites.

The pricing is its own signal. At $3 per million input tokens and $15 per million output, K3 is the most expensive Chinese model ever released, a deliberate break from the discount strategy of DeepSeek and Z.AI, but still well under a third of Fable 5's output pricing. That is Sonnet-level pricing for near-Fable intelligence.

One practical note: K3 currently runs only at maximum thinking effort, and it thinks a lot. A simple SVG task burned over 16,000 output tokens, and in my sessions it ran 2-3x slower than Claude. Budget for both.

But the open weights are the real story. Any provider can serve it, any company can host it inside their own walls, any researcher can build on it. A frontier-level model you can run in your own infrastructure was not on the menu six months ago, and for anyone with sovereignty requirements that changes the conversation. Until July 27, everything routes through Chinese servers; whether that matters depends on what you are sending it.

Self-hosting will not be casual either: Moonshot recommends 64 or more accelerators, so local deployment belongs to well-funded teams for now, while everyone else fine-tunes through cloud providers. Moonshot has already contributed a prefix-caching implementation to vLLM so serving works on day one.

K3 does not change who has the best model. Fable 5 and Sol remain ahead, and the polish gap is real in everyday use. What it changes is the shelf life of a lead. Dario Amodei and many others expected Chinese labs to stay several months behind the US frontier; for open models, that cushion is now roughly zero, and it shrank while Western labs were gating their strongest systems behind approval tiers and navigating export restrictions. Whether restriction or distribution wins is, I think, genuinely undecided, and Beijing has visibly picked its side, with Xi Jinping championing open-source AI at the same Shanghai conference where K3 debuted.

For anyone running AI in an organization, the practical implication is quieter than the market reaction: the frontier is about to be something you can hold, and that reshapes procurement, sovereignty conversations, and fine-tuning roadmaps more than any single benchmark.

Exciting times!

There are three ways to try K3 right now.