In partnership with
IN THIS ISSUE
Welcome back to The AI Field.
The frontier used to have a return address. This week it doesn't. The best coding model on the leaderboards came from a lab nobody could name, OpenAI published numbers showing its own chip beating Nvidia's, and Perplexity moved a working agent onto a box that sits on your desk.
- Z.ai turns out to be the lab behind the stealth model that topped the charts
- OpenAI publishes its first Jalapeño benchmarks against Nvidia Blackwell
- Perplexity ships a Computer agent that runs locally and skips the cloud
- Hugging Face is reportedly weighing bids above $13 billion
Read time: 5 minutes
200 Ways To Make Money With AI
Ready to transform artificial intelligence from a buzzword into your personal revenue generator?
HubSpot’s groundbreaking guide "200+ AI-Powered Income Ideas" is your gateway to financial innovation in the digital age.
Inside you'll discover:
A curated collection of 200+ profitable opportunities spanning content creation, e-commerce, gaming, and emerging digital markets, each vetted for real-world potential
Step-by-step implementation guides designed for beginners, making AI accessible regardless of your technical background
Cutting-edge strategies aligned with current market trends, ensuring your ventures stay ahead of the curve
Download your guide today and unlock a future where artificial intelligence powers your success. Your next income stream is waiting.
RECENT UPDATES
01 · OPEN MODELS
For a week a model called Ox Alpha sat at the top of OpenRouter and OpenCode with no lab attached to it. On Wednesday Z.ai confirmed it. Ox Alpha is the newest model in the GLM series, and the weights go out openly the same night.
It's built for coding, long agentic runs, and production work, with a context window just over a million tokens and text, image and video input. It ended DeepSeek's 56-day run at the top of OpenCode, and it's free to use on OpenRouter right now.
Independent scoring is still thin. The early benchmark posts are small samples run by individuals, so treat them as directional until the weights are in more hands.
The AI Field: The interesting part isn't that a Chinese lab is at the top. It's that it got there anonymously. For a week Ox Alpha was judged on output alone, with no brand and no price attached, and it won. Once the weights are free, the question for a paying customer stops being which lab is best and starts being what the premium actually buys.
TRY THIS TODAY
Test whether a cheaper model can do one of your jobs
Three stories in this issue point the same way: capable models are getting cheap. Here's a 15-minute way to check whether that's true for your work.
1. Pick one task you send to a frontier model over and over, such as summarizing calls, drafting replies, or cleaning data.
2. Open a fresh chat with the model you use now and paste this prompt:
I want to test whether a cheaper model can replace my current one for a repeated task. The task is [TASK]. Here are three real examples of my inputs and the outputs I accepted: [PASTE]. Ask up to three short questions first if something important is missing. Then write five pass-or-fail checks describing what a good output must contain, based only on the examples I gave you. Do not invent standards I did not show you. Give me a plain checklist I can score by hand. End with the single failure that should stop me from switching.
3. Run ten real inputs through the cheaper model, score them against the checklist, and only switch if that one failure never shows up.
02 · AI HARDWARE
OpenAI published the first performance numbers for Jalapeño, the inference chip it designed with Broadcom. Across GPT-OSS 120B, DeepSeek R1 670B and Kimi K2.5 1T it reports 1.5 to 1.9 times more work per watt at peak throughput, 1.7 to 3.6 times lower end-to-end latency, and 2.1 to 4.1 times higher performance on interactive workloads.
On the biggest model tested, that came out as 18,195 mixed tokens per second per kilowatt against 11,862, and 1.56 seconds of end-to-end latency against 5.31. The chip is rated at 700 watts and measured at or below 550.
Two things to hold onto. The comparison is against Nvidia Blackwell, which won't be Nvidia's newest part by the time Jalapeño is everywhere. And OpenAI hardware chief Richard Ho says deployment starts at the end of 2026 in very small volumes, with real scale in 2027.
The AI Field: These are OpenAI's numbers on OpenAI's silicon, so read them as a claim rather than a verdict. The claim is still the point. Nvidia doesn't lose a customer here, it loses the assumption that there was no alternative, and that shows up in contract pricing long before it shows up in market share.
👉 read more here
03 · LOCAL AI
Perplexity launched Portable Computer, a local-first version of its Computer agent that runs entirely on your machine and only escalates to the cloud when a task needs it. It runs on NVIDIA's DGX Spark, the Grace Blackwell GB10 desktop with a 20-core Arm CPU and 128GB of unified memory, with RTX GPU PCs coming.
Locally it runs Qwen 3.8 27B and PPLX 27B, Perplexity's own post-trained Qwen variant, with NVIDIA Nemotron 3.5 Lightning 30B on the way. It's live for Pro and Max subscribers, Linux first and Windows later, installed in one click. Work done locally doesn't burn credits.
The AI Field: Local models stopped being a hobby the moment one could finish a real job. The pitch isn't only privacy. It's that the marginal cost of running your own agent drops close to zero once the hardware is paid for, and that math is going to decide a lot of budgets next year.
04 · INDUSTRY
Business Insider reported over the weekend that Hugging Face is in talks to be acquired at $13 billion or more and is working with banks to evaluate bids. No buyer has been named and nothing is signed.
For scale, its last round in 2023 valued it at $4.5 billion post-money, led by Salesforce Ventures. It turned down a $500 million investment from Nvidia earlier this year that would have valued it at $7 billion. CEO Clem Delangue has said the company is close to profitability and has talked about a long-term responsibility to its community.
The AI Field: Hugging Face is where open weights live, including the Z.ai weights landing tonight. Whoever owns that index owns the default distribution channel for open AI, which is worth considerably more than the revenue on the books. Watch who bids, not where the number lands.
THE AI TOOLBOX
FEATURED TOOL
What it does: Describe an app, game, dashboard or site in the chat and Grok returns a live working version you can publish to a grok.me address, remix, or export to GitHub.
Best for: Founders and operators who want a prototype, internal dashboard, or landing page without opening an editor.
Pricing: Now on every plan across web, iOS and Android. It was limited to SuperGrok Heavy during the early beta.
Worth trying: Build one small internal tool you would otherwise ask an engineer for, then read the exported code before anything real depends on it.
Also on our radar:
Keenable: A web index built for agents instead of people, with more than 100 billion documents behind an API. It raised a $26M seed led by Accel.
QueryStory: Enterprise data answers that show the SQL and a confidence score, so a person can check the work before acting on it.
Instinct: A capable personal agent in private beta. Read the terms before you connect an inbox, they claim a perpetual license to your material.
QUICK HITS
Claude now remembers across chat and Cowork: Anthropic merged memory between the two on Free, Pro and Max, so a project you briefed in chat carries over without a re-brief.
Zuckerberg's plan to replace Meta staff with AI fell apart: Reuters reports Meta modeled cuts of up to 60% on some teams, then pulled back after internal data showed the agents underdelivering and a 40% rise in technical incidents tied to AI-generated code.
OpenAI asks California to strengthen SB 53: The company that once opposed the bill now wants monitoring during training and evaluation, plus tougher security requirements, written into it.
YOUR TURN
What did you think of today’s edition?
Until next time!
Olle
Founder of The AI Field