ShellDrop iOS

Worked with Fable 5 on a security audit of ShellDrop today. It used Glasswing - Cloudflare’s multi-agent security-review pipeline. Agents were 8 hunters, 3 verifiers. The architecture strengths it confirm: •SSH passwords + private keys live in the iOS Keychain, device-only. Never synced to iCloud, excluded from backups. •Secure Enclave-backed keys. The private key is generated inside the hardware and never leaves it; the app never holds the bytes. •No analytics or telemetry, no third-party SDKs. Nothing phones home. •Nothing sensitive is logged. No passwords, hostnames, or terminal output written to logs. •TOFU host-key verification. Pins each server’s key on first connect and blocks the session if it changes (MITM protection). •Only public key + metadata sync. Private key material never touches CloudKit; connection sync carries no credentials. •Encrypted/malformed key import is rejected safely. The key parser is bounds-checked, no crash-on-hostile-input. •Keys generated with the system CSPRNG (CryptoKit), and there’s no private-key export path (only the public key can be copied). •App Store privacy label: “Data Not Collected.” True, the app transmits nothing.

@LLMJunky Spot on about the hype. It is so boring now

Yep. Kimi K3 is a very good model. Better than Fable and 5.6 Sol is cope, but one can be forgiven for saying it given they had no benchmarks. And that is the trend around here. Overhyping. It is what it is. Anyway, this is an incredible model regardless. DeepSWE: 67.5 (3rd) Terminal Bench: 88.3 (wow!!) (2nd, but barely) FrontierSWE - 81.2 (comfy second) I do think its safe to say, based on the early testing, and these results, that Kimi has achieved near-parity, with our existing SOTA models. It wouldn't surprise me either if it was one of the very best models in UI design, spatial awareness, and writing. A job well done @Kimi_Moonshot - thank you for sharing the weights to this incredible model. Now I just need to find a way to get 3.5TB of VRAM.

@105Hasnain @maxedapps How do you find the CLI compared to Claude, Grok, Hermes, etc.?

@maxedapps I’ve been using Kimi Code CLI. I compared a couple of results using the same prompts with Fable and Sol, and the results were pretty impressive, its actually close. But I am still going to test more.

So are you all using Kimi K3 via their API?

@LLMJunky That's not the only issue, the $19 plan doesn't give you the 1M context window for K3... You're on 256k

The rate limits on Kimi Code on the $19 plan are really bad imo. I only got through about ~85% of building a browser game before hitting the 5hr quota. This burned 20% of my weekly, so I only get 4 more quotas for the week. That was one prompt, and it didnt even finish. This is not a large project. Fable one shot this for me on the $20 plan without running out at all.

@deredleritt3r Could you clarify the areas which you think are materially worse?

A few thoughts on Kimi K3: - The Moonshot team is absolutely cracked. Some of you might remember that K1.5 was released on January 20, 2025 - i.e., on the same day as DeepSeek R1, and it was ~just as good as R1. IIRC, the model wasn't open-source, which is why no one ever talks about it and DeepSeek got all the glory. - Do K3's capabilities generalize outside of coding, or is it coding-benchmaxxed? In my initial testing, it's materially worse than GPT-5.6 Sol and Fable 5 for certain use cases that do not involve coding. - I will not tire of saying this: the publicly available models in July 2026 *do not matter*. It doesn't matter whether China is 2 months behind the frontier, or 4 months, or 8 months. The only thing that matters is the race to RSI. What do we need for RSI? IMO, two things: (i) a model with excellent research taste; and (ii) tons of compute. Does excellent research taste develop on its own from increasingly strong coding abilities? I'm skeptical. I think it's a much more general ability that is more a function of overall model intelligence than just proficiency at some skill (like coding). - All that said, my estimate for how far behind China is has not changed at all with the release of K3 (IMO, much farther behind than the timeline thinks today). - This model's cyber capabilities are currently unknown, and this is an area worth watching. The few weeks following the release of K3's weights might be interesting times for defenders (or, then again, maybe not). - The geopolitical consequences of the release of this model are also worth watching - both in the U.S. and in China.

the Open-Source Capabilities Flippening cometh

@doodlestein @amar_patel How would you describe the speed of which it works? Faster than Opus 4.8? Not as fast as Grok 4.5?

@amar_patel Yes, their $200/month subscription

OK, let's see what this sucker can do on the ~1.5mb markdown plan for my new FrankenGraphDB project. I'll post the results below so you can follow along. This plan has already been exhaustively reviewed by both Fable xhigh and Sol Ultra, so the low-hanging fruit is gone now:

Interesting point here

@alsanaree @scaling01 But have you tried K3? 2.8T … that’s a big boy

@scaling01 open source models have that “yeah you can use me for fun and then get back to your 200 dollar codex/claude sub” feeling it’s kinda like xiaomi or HUAWEI managers using iPhones back in 2015 Chinese models have a long way to go

from Moonshot's own blog post: "Despite being a highly competitive model overall, K3 nonetheless exhibits a noticeable gap in user experience compared with Claude Fable 5 and GPT 5.6 Sol"

@scaling01 But what does "user experience” mean?

@thdxr I haven't tried it personally just yet but the big thing that it's different here is the $2.8 trillion parameter training data size ...

@AlexFinn Available today with only 3 x $20k mac studios and 512gb of ram, why isn’t everyone doing this.

I was wrong. I said we were a year away from Fable 5 on our desk. That day is today An open model BETTER than Fable 5 in some benchmarks just dropped Better than ChatGPT 5.6 on FrontierSWE. Better than Fable 5 on Automation Bench This fundamentally changes the AI race forever If people can start running Fable 5 on their desk, unlimited and for free, they are not going to pay thousands for subscriptions Now let's be clear: will the hardware you need to run Kimi K3 be obtainable for most? No it won't It will require multiple very expensive Nvidia chips or a bunch of Mac Studio 512gbs But look at the slope, not the Y intercept. Over the last year local AI has become SIGNIFICANTLY more efficient and required way less compute The smartest brains in the world are all attacking the compute problem as hard as they can This is another step in that direction. And within a year or two, you'll be able to run this on a Mac Mini Local AI has arrived, and it's not going anywhere

@theo Anyone tried the native Kimi TUI?

What are the best harnesses to use with Kimi K3 right now? It heard it's pretty good at orchestration, where's the best place to see that?

@arena @Kimi_Moonshot Curious about what this says about the models general intelligence? Most of these exams means nothing when it comes to thinking, reasoning and and real decision making.

Big news: Kimi-K3 by @Kimi_Moonshot is now #1 in the Frontend Code Arena with 1679 pts, surpassing Claude Fable 5. This is a 17-place jump from Kimi-k2.6 (#18 -> #1). In Frontend, Kimi-K3 ranked #1 in 6 of 7 domains: Brand & Marketing, Reference-Based Design, Data & Analytics, Consumer Product, Simulations, and Content Creation Tools, landing #2 only in Gaming behind Fable 5. The full model weights will be released by July 27. Congrats to the @Kimi_Moonshot team on this major milestone!

Kimi K3 isn’t cheap.

@thsottiaux @AndrewCurran_ do you guys ever share how you calculate this active users number? people are saying it looks like you're blurring the lines there for the sake of hype metrics which are empty numbers in reality because of the merging of your two major apps

Another reset for our Codex and ChatGPT Work users. Actually hit 9M active users way earlier today, but then got distracted by the approximately millions of things the team is doing to keep the systems up and reliable. Should have that sweet 100% weekly usage limit back in a few minutes. Go be your productive self and close twitter. Shoo!

What's the best general intelligence test or exam that these ai models take, one that actually reflects actual strategic reasoning and thinking, real life intelligence. The one where fable is number one, opus 4.8 is number two, and then everything else is below.

the funny thing about this is yes it's 50x cheaper technically but it's not accessible at 50 times less the price of anything... in the real world the 50x metric doesn't translate. Even using opencode go with their $10 per month plan, you'll blast through the kimi k3 usage in 6 hours for the entire month. then you'll be using api credits and you'll be paying $100 per day

Kimi K3 is about to have America rethinking its life choices tomorrow. What do you do with a Chinese model that’s approaching SoTA, but 50x cheaper? Oh, yea. And it’s open-source.

@Kimi_Moonshot Wen 3.0

Thats a cool product. There will be more.

@MatthewBerman But if you want to stream thought directly to an agent, a great VTT service is your best friend

I think more clearly when I type than when I dictate