The Local Pass — Kimi K3, passkeys, and harnesses becoming claws

00:00:04 Wednesday morning, Moonshot released Kimi K3. It's a 2 point 8 trillion parameter model with a million token context window and native image inputs alongside text. That last part is notable — most models treat images as an afterthought bolted onto text architecture.

00:00:23 K3 reasons across text, audio, and video as a single unified pipeline. The specs alone make it unusual in the open-weight world. Until now only a handful of open models have even reached the trillion parameter range, starting with Kimi K2 last summer. DeepSeek V4 Pro came out in April at 1 point 6 trillion parameters, followed by Xiaomi's MIMO v2 point 5 Pro at a trillion, and Thinking Machine's Inkling just under the same mark.

00:00:55 GLM 5 point 2 from Z.AI took all that bluster after hitting Fable 5, but sat at only 744 billion. Moonshot's lead researcher Ginnyu Yang posted about why they could ship K3 while others couldn't back when she left academia for industry.

00:01:13 She described four patterns at other labs: arrogance, the belief that the AI war is over and they won; restlessness from young labs rushing to catch the frontier or pivot away from competition; fear from strong second-tier teams who can't quite bring themselves to aim for number one; and misalignment where everyone's optimizing for their own credit but nobody actually cares whether the company reaches AGI.

00:01:43 Moonshot was different. Over many conversations with the founders, she said the same thing came through every time: a raw kind of hunger. She ended up joining them, and they got K3 out the door.

00:01:58 The benchmark numbers are strong on paper. On DeepEval's coding eval it hit 67 point 5, putting it 8 point 5 ahead of Opus 4 point 8 and half a point ahead of GPT-5 point 5. Artificial Analysis backed that up with an intelligence index of 57. It sits third overall, three points off Fable 5 and two behind GPT-5 point 6 Soul.

00:02:22 On Terminal Bench 2 point 1 K3 hit 88 point 3 — a half point behind Soul and a few points ahead of Fable 5 across the board. On VAL AI's index it did even better, taking second place overall and surpassing GPT-5 point 6 Soul with a 20 percentage point improvement over its predecessor in less than three months.

00:02:45 It's also the first open-weight model of its size. Early testing tells a different story than the benchmark grids do. One user gave it a simple hover color bug in a TUI application. GPT-5 point 6 Soul found and fixed it with 30 cents of spend. K3 got up to one dollar and started reading their database before being interrupted.

00:03:09 Another user ran it through a coding task where the model consumed over 13 thousand reasoning tokens to produce just three thousand four hundred output tokens, which drove up the bill. Multiple users reported reasoning loops that burned tokens on simpler tasks.

00:03:28 Pricing tells the same story. Blended input-output costs sit at five dollars and forty cents per million tokens. That's less than Opus 4 point 8 at nine dollars and GPT-5 point 5 at ten, but a lot more than DeepSeek V4 Pro at four cents per task. Ryan Fedrick pointed out the hardware reality: to hold a 2 point 8 trillion parameter model in silicon you would need roughly equivalent to 44 Mac Studios or 15 Blackwells — an NVL72 rack costing hundreds of thousands of dollars.

00:04:03 The model sits alongside frontier proprietary models in cost and difficulty to run, not with the cheap open-weight alternatives. On Baidu's LiveBench K3 topped open-source but still fell short of Soul and Fable. Users report it spins a lot on Opus-class problems, runs slower than frontier models, and gets caught in reasoning loops.

00:04:28 The model has real strengths, though — particularly in frontend code generation, where Arena.ai ranked it first across six of seven domains. Its native multimodal architecture lets it reason across text, audio, and video in ways many Western models still bolt onto text.

00:04:48 That architectural difference matters: different training data, procedures, and attention mechanisms let K3 blend well with Soul and Fable to surface problems both of them missed.

00:05:01 Nikita Bier posted a thread this morning about passkeys that resonated across the platform with 213 replies and over three thousand likes. His thesis is simple: passkeys were invented by security engineers with zero understanding of the consumer brain. We had no way to argue its merits since no one understood what they are, he wrote.

00:05:25 As a result we are now all asked to produce magic fairy dust to login to apps. It may be on your phone in your browser in your operating system in your body who knows. But now you're safe.

00:05:38 The replies painted a picture of real user confusion. Jason Levin tweeted that he runs a tech company and has no idea what a passkey is and at this point is too afraid to ask. Edwin Hayward flagged a concrete problem that keeps coming up: when your only passkey device disappears, you're locked out.

00:05:59 Passwords can be written down somewhere; two-factor apps can live on multiple devices. A single-device passkey maximizes your chance of losing access while offering robustness only when everything goes perfectly. Thaddée Tyl, who goes by espadrine, put it more precisely: passkeys are outstanding cryptographically but the design complexity leaked to the UX.

00:06:24 He referenced a specific incident where passkeys on X stopped working because of domain protocol differences between http and https. The core tension comes down to trade-offs between cryptographic guarantees and everyday usability. Security engineers optimized for worst case attack scenarios — phishing resistance, no server-side secrets to steal — at the expense of everyday usability.

00:06:52 That gap between what they built and what people actually experience in a browser login flow is where things break. Dom Hofmann tried to clarify the distinction: instead of a word you type, it's a key you carry. Most people use their phone as that key, though you can also use a laptop or other devices.

00:07:13 Words are easy to steal; keys are harder. The problem isn't the cryptographic model — it's that nobody told users what they were actually carrying. It's a recurring pattern in infrastructure design. When the people building a security layer haven't used their own product, they optimize for the attack surface and forget the human on the other end.

00:07:38 Sam Bhagwat, co-founder and CEO of Mastra, framed the current moment at AI Engineer last night with a title that stuck: every harness will become a claw. He's describing a shift most developers haven't fully registered yet. The old model was straightforward — you talk to an agent in your coding environment, it talks back.

00:07:59 The new one is more like what he calls claws: always-on systems with durability and doggedness that can run for hours or days, not minutes. The key affordances are coming through now. Planning mode lets agents decompose tasks before executing. Parallel sub-agents fan out multiple work streams simultaneously.

00:08:20 Skills get created dynamically rather than predefined. Session-long tool approval chains replace per-call authorization. Agents autocompact when context fills up and resume from where they left off after connection drops. What's actually changing here isn't about the models getting smarter.

00:08:39 It's about the framework layer gaining persistence — agents that live on Slack alongside your team, mobile apps that tunnel into local machines, and cloud sandboxes that run dozens of tasks simultaneously. The harness stops being a tool you invoke and starts being infrastructure you build around.

00:08:59 Bhagwat borrowed an analogy from someone actively using agent frameworks in production: durability is what separates an agent from a harness. It's the capacity to persist through failures, autocompact when context fills, and run background tasks without human oversight.

00:09:17 They're plumbing problems, but they also mark the difference between something that works in a demo and something you ship. The harness era isn't about agents replacing humans. It's about systems holding state long enough to actually finish the work.

00:09:34 Wednesday boiled down to three layers of the same stack. Models keep getting bigger and closing benchmarks, with Kimi K3 taking the crown for open-weight models despite being expensive, slow, and still not matching frontier ones on real codebases. Agent frameworks are moving from local coding assistants into always-on infrastructure with persistence and parallelism.

00:09:56 That transition from agent to harness is about durability, not intelligence. And the user-facing layer — authentication identity — remains frustratingly misaligned between what engineers built and how people actually use it. Everyone agrees on the model in theory.

00:10:12 Nobody has solved the problem in practice. The passkeys thread was the sharpest thing I read today. It's a reminder that the gap between security engineering and consumer experience is where most of the friction lives in this industry.