Hey from Tobias,
Kimi K3 is the biggest open model I have seen: 2.8 trillion parameters, a 1 million token context window, and open weights landing July 27. But the thing I kept coming back to this week was not the size. It was how much of great use cases are about doing less, not more.
Four things below: K3, a prompt that turns AI into an idea machine, why local AI is not about saving tokens, and the framework tax nobody talks about.
Kimi K3 Is the Open Model to Watch
The specs are large even by 2026 standards. Kimi K3 runs 2.8 trillion parameters, reads a full 1 million tokens of context, takes text and images, and is built for long, multi-step coding jobs. A new trick called Kimi Delta Attention makes it decode up to 6.3x faster at that million-token scale, so the giant context window is actually usable and not just a spec sheet number.
Early results back it up. It sits at number one on the Frontend Code Arena and holds its own against Claude Opus 4.8 and GPT-5.6 on several coding and agent tests. The honest catch: some users find it slower and more verbose in its reasoning. Open weights are planned for July 27, and once they drop, more providers will start hosting it.
If you want a cheaper Claude Code alternative right now, three subscriptions already work with OpenClaw and Hermes Agent:
• GLM-5.2. Strong coding model through the zcode plan on z.ai.
• GPT-5.6. Included with a normal ChatGPT subscription.
• Grok 4.5. Available on SuperGrok or higher.
AI Has Zero Creativity? The Data Disagrees
Take-Two CEO Strauss Zelnick, whose company owns Rockstar and GTA, said that no AI model can be creative because it is data-driven. A Wharton study says otherwise. Ethan Mollick and his colleagues put GPT-4 against 200 elite MBA students, all generating product ideas for college students under $50.
The AI produced 200 ideas in about 15 minutes. On blind consumer ratings, 35 of the top 40 ideas came from the AI, and purchase intent ran 47% for AI ideas against 40% for the humans. Mollick calls large language models connection machines. They do not invent from nothing. They recombine distant, unrelated ideas, the same way the Wright brothers connected bicycle mechanics with bird flight.
The way to use that: force three unrelated things into one prompt and make the model connect them. Here is one you can paste right now.
1. Silent discos (wireless headphone parties)
2. Ancient Roman thermae (public baths as social and health hubs)
3. U.S. Patent US10869621B2 (piezoelectric flooring that harvests energy from footsteps)
Requirements: full business name and one-line pitch, target customer, how it works, revenue streams, why the three elements make it special, and the first three things to prototype.
I ran it and got Thermae Pulse: a Roman-style bathhouse where guests wear silent-disco headphones and the whole floor harvests energy from every step and dance. Entry, private buyouts, energy credits, merch. Was it perfect? No. Your job is to kill 80% of what it gives you and keep the one idea that makes you stop. The AI is the idea machine. You are still the taste.
Local AI Is Not About Saving on API Bills
The common pitch for local AI is running a frontier model on your own hardware to dodge API costs. That is not the interesting part. The real win is small, focused models on cheap hardware you already own. Your data never leaves the machine. It works offline. No vendor can read it, no rate limits, no surprise bills.
I built three production apps this way: a CRM that imports and enriches contacts on its own, an ERP you run entirely by voice or text, and a meeting recorder that transcribes and summarizes everything on-device. None of them try to be as smart as Fable 5 or GPT-5.6. They do not need to. They need to be reliable, private, and always there. For most real workflows, privacy and control beat raw intelligence.
Most Agent Frameworks Are Taxing You
Here is my hot take for the week. Most AI agent frameworks quietly tax you on tokens, bugs, and reliability. They add system prompts, handoffs, extra loops, and retry logic that fires more often than you would expect. What could be one clean run in plain Python turns into three to five calls plus retries, and the token bill can land at multiples of the simple version for the exact same result.
Models also know Python far better than any framework's API, because they were trained on mountains of it. Stay in native Python and the model hallucinates less and writes safer code. Frameworks excel on genuinely complex multi-agent work. But a huge share of agent projects are overengineered from day one. Keep the surface area small and you stop chasing a new framework every few months.
Do less. Ship more.
Tobias
