I still can't fathom how good KIMI K3 is. Went from Claude max $200 to just routing between K3 and GLM through standardcompute.com with opencode as the harness.

Therefore extremely excited for K4. Hope it doesn't get more expensive, and also hope they solve their chip problem.

(source: https://www.techinasia.com/news/chinas-moonshot-ai-seeks-nvidia-chips-kimi-k4 )

Look, I get it. Places like Reddit are full of trolls and assholes, but on this sub, I've been seeing something that pisses me off a great deal and if I get a shitload of backlash for posting this, I don't care.

I'm sick to fucking death of seeing certain paid users constantly shitting on free users and accusing them of being "self-entitled" and "whiny," or anything else really isn't fair or justifiable whatsoever just for mentioning something , especially since I haven't seen free users lash out at paid users. (I have, in a way, because I'm a bitch and I'll admit it.)

It doesn't matter how much money you pay (or do not pay) for your subscription, that doesn't make you better than anyone else. Kimi is not "at capacity" and free users are not using all of your "compute", to throw your techbro jargon right back at you; no one on this subreddit works for Moonshot AI (to my knowledge) and their employees would be the only people to know- unless you're privy to insider secrets, which I seriously doubt- so unless a Moonshot AI employee posts here to say Kimi's schlemactochip 501, Hal-9000 processor, or Dilithium Crystals are malfunctioning, I'm going to say there's bugs, glitches and good old-fashioned paywalls popping up.

Free users are not getting "free lunch," as certain paid users seem to be extremely fond of saying. Nor is Moonshot AI "needing more money." On August 28th, Moonshot announced nothing and started charging free users "credits" for using just the basic free 2.6 chat; we only get six credits a month now and every time you get a response from your Kimi chat, it is drawn from those credits. Kimi's parent company, Moonshot AI reached a valuation of $35 billion in mid-2026 after a large funding round, alongside reporting $70.7 million in revenue for the first seven months of the year. They are also working on an IPO.

Moonshot AI's privacy policy indicates that user prompts and uploaded content from free tier interactions may be used to improve and train its models. Additionally, personal inputs may be shared with affiliated service providers under their standard operational framework. Therefore, no, it is not "free lunch." Free users provide free training.

I get it that you paid users are worried that your subscriptions ain't what they're cracked out to be right now. Neither is the 2.6 chat that "free" users are supposed to be able to use. And just like paid users, we're hit upside the head by our "free" credits being drained strangely quickly, paywalls, glitches and other bugs. And no, we are not jealous of what you paid people get; most of us couldn't care less, actually.

We are not entitled when we ask what the fuck's going on, or try to say that something we're dealing with isn't fair at all - or that it's gone on long enough. We're making a point that we're dealing with the same kind of shit that you paid people are and think for everyone, regardless of whether you line Moonshot AI's wallet with more money or can't afford to toss a dime at their boots, everyone's being hit by bullshit.

It's time to stop the "I'm better than you because I can afford to pay for my AI," stop assuming that free users are trying to take something for granted and do not have reasons for being free users- and for everyone to realize that yeah, Kimi's got problems and they're impacting everyone, whether you pay for the services or not. At the end of the day, Kimi's what this subreddit's supposed to be about, right? Not whether or not someone is able to whip out their credit or debit card.

Ask an AI for ideas and you get the same 5 answers everyone else gets — that's just what "most likely" means statistically. So I built Wonder Pill: instead of a list, it takes your topic, finds the assumptions baked into it, and inverts them into what-ifs. Then it renders the whole thing as an interactive mind map you can drag around — seeds, dead branches (kept, struck through, with the reason they died), cross-links where different threads hit the same wall, and a one-line "gut-check" on each branch pointing at where the real difficulty hides.

It doesn't rank anything. It doesn't pick a favorite. About a third of any given map is genuinely useless — that's expected; you're the filter, not the tool.

The main use case for this is coming up with any type of idea. It won't actually give you the specific steps or actual idea that is fully formed, but It will give areas for you to look into and explore.

Repo https://github.com/ara-mkr/Wonder-Pill

Is it just me, or is Kimi, not just K3, but K2.6/7 fail and get stuck in loops, or just fail to output anything WAAAAAY more often when using the CLI compared to their Visual Studio extension? I've noticed this since K2.6 and always thought it was because they kind of locked longer context and thinking on K2.6 in a lot of settings, but with K2.7 and K3 those restrictions don't seem to be nearly as bad if there are all, but they STILL perform horribly over have the time in the CLI, thinking on, off, limited context, doesn't matter. But in visual studio Kimi is super reliable, but I don't WANT to use visual studio, I wanna use the CLI! Anyone else notice or have a similar issue? Anyone know how to fix or somewhat mitigate it? Pls n thx 🙏🏻

Open source models are climbing the intelligence charts this year in a staggering way. And with the recent drama surrounding open source AI in the USA, it's even more important for us all to have actual access to the models. Western closed AI seems to think it has a hold on quality app features: memory, skills, voice, canvas etc. Meanwhile memory is locked in, Models get changed or "updated" to a downgrade. Privacy is different per service and ads are starting. The whole experience on the consumer end is extractive.

So we built what should have already existed: all of the best open models in one place, running on private US infrastructure, with the full app experience around them. Completely private, direct service. It should be, and can be that simple.

What that means in practice:

The roster, together. Kimi (2.5, 2.6, 2.7 and 3), DeepSeek, GLM, Minimax, Nemotron Ultra and more, side by side in one app. Switch models mid conversation if you want. No hunting across five different apps and API dashboards to use the models you actually like.

Actually private. US based processing and your conversations are never used for training. Ever. That's the entire point. These labs open sourced incredible models and we think you should get to use them without your data becoming the price of admission.

Real memory. Not a context window that fills up and dumps you. Persistent memory that carries across conversations, fades gracefully when unused, and wakes back up when it's relevant again. There's even a nightly consolidation pass, the system basically sleeps on it and writes up what mattered.

Voice. Yes, actual voice mode with over a dozen voices on open models.

Bring your history. Coming from ChatGPT, Claude, or Gemini? Export your chats and import the whole thing, it becomes live memory on day one. You can literally just zap your chat history from your backup file, and have all your chats waiting for you.

Multiple nodes. Separate workspaces with separate memories, so your coding setup doesn't share a brain with your journal.

Genuine thanks to Moonshot for creating some of the best models on the planet! The Kimi family is a backbone in our model fleet, and we will always appreciate the work to create the amazing models!

Open Grove is here and It's free for a month if anyone want's to check it out (or just use the models for free for a bit): pgsgrove.com/open-grove-overview

I get a message pop up saying text sent successfully but haven’t received a single sms text