You can try to use LibreOffice all you want, until you receive a form to fill where LibreOffice doesn't display checkboxes and recipient won't accept "x", because their automated software only uses MS Office and won't properly attribute it to questions.
The bad news is that the original v4 flash was 284B, which was large but still somewhat reasonable for running locally. This one is 552B so almost twice that, so the huge gains in benchmark scores make sense - it's not really flash anymore, imo.
I've no idea about actual performance vs benchmaxxing, though deepseek was fairly trustworthy as far as Chinese models go. If that holds (and if it doesn't think forever, as deepseek 4 sometimes did) it's probably the newest king of the hill amongst open weights models.
It does include vision, and they do something funky with KV cache so it's very efficient: "[...] these designs reduce the global KV cache footprint to 890 bytes per token — roughly 1/4 of DeepSeek-V4-Flash". I do appreciate the high focus on efficiency, but at this point we sure could use a flash-flash version.
@edit: I couldn't make sense what the actual parameter count is, with the addition of Engram memory. To my understanding the 4.1 flash is 552B parameters you want in vram or ram, out of which ~16B is active (8B for prefill). It also includes additional 196B Engram memory which you can put on an SSD. I think.
Assuming that's correct 256 GB memory is insufficient to even load the model at q4 - you'd be 1GB short, assuming you can fill it to 100% (so no mac). You'd also want some for kv cache of course. A 256 GB desktop with some extra VRAM from GPU could run it, but normal consumer boards get real slow once you fill 4 slots so you'll probably want quad channel which is Threadripper or above territory.
Id love an ELI5 for PLE. Im trying to work it into my back of the napikin math for compute vs memory bandwidth limitations on tok/s in PP vs TG work.
My attempt at a simplification of this article on it https://sebastianraschka.com/llm-architecture-gallery/per-la... into a couple of sentences is that they are linear embeddings of the input token space projected per layer, which are then gated by the transformer outputs per layer.
This would mean that the only one set of weights for the ple path needs to be pumped across the memory bandwidth as they are the same linear weights for all layers?
Sheit, maybe im trying to simplify something that i need to look at in detail. but id love to leverage others understanding if possible
While he avoids using the actual PLE acronym, he does actually describe the concept quite well. I think you may enjoy this video. Specifically around 5 minutes into the video is the part you're looking for.
Flash is about speed ... Flash models are supposed to be fast, way faster then their big brothers that are "better" but way slower.
Its just that up to now, getting more speed involved cutting back on the parameter count, what ended up making the Flash models more "dumber" in exchange for speed.
What we see with DS v4.1 Flash, is that DeepSeek has found a way to make a Flash model, that is 2x a 2.5x faster then the older Flash version, while increasing the intelligence (more parameters). To the point that it goes past Kimi K3 and GLM 5.3 in most tests, with a blazing 250 to 400t/s.
AND its also priced as a Flash model (they even reduced the price back to almost old v4 Flash price), despite it now rivaling those 10x to 30x more expensive competitors.
The issue that people can not fit it into local setups, is not how companies design their models. They design it for their own needs. A old flash needed less parameters to be fast, and local users had the benefit of it fitting in 256GB memory.
Companies who run locally, are perfectly able to buy a few H200/B200 and get a setup that run a model that almost rivals Opus 5.0 in their office. How to say this without getting downvoted. People get way too fired up if a model does not fit, despite that they can still run the old v4.0, qwen 27b, 35b, 3.8 Next and other models. The fact that these models are being released for free, is already amazing by itself. I am still waiting to see what Anthropic and OpenAI and Google are releasing for free... O wait ... ;0
> Companies who run locally, are perfectly able to buy a few H200/B200 and get a setup that run a model that almost rivals Opus 5.0 in their office.
I agree with your broader point about Flash being about speed not total model size, but I think we should also point out that H200/B200's are seriously overkill for the "run a model in your office" scenario. That sort of hardware is optimized (in a roofline analysis sense) for running hundreds of concurrent sessions on a 24/7 basis. You're severely overpaying for your VRAM in basically any typical local-inference scenario, you should most likely be buying gear based on LPDDR and Flash memory instead which will slash your cost by orders of magnitude.
> I think we should also point out that H200/B200's are seriously overkill for
I simply mention what came to mind ;)
A quad 6000 with 96GB, can run this model at NVFP4. That is 60.000 Euro for the GPUs and lets be generous with another 20.000 for the rest of the system. The price of a single developer for a year.
I highly doubt that it goes past Kimi K3 in actual practice. GLM 5.3 claimed the same, but in practice, K3 is so damn knowledgeable and I suspect due to it's massive size.
V4 Flash also was released as mostly FP4, but this one is FP8 (?).
160GB vs 510GB.
Original Flash good fit for dual Spark / Strix Halo machines. This one would require third party quants and even then 4 machines.
Edit: Most of added weights/size are Engrams?
> Overall, DeepSeek-V4.1-Flash has 552B backbone parameters and 196B Engram parameters, activating 8B parameters per token during prefill and 16B during decode.
Those can stay on SSD. So I guess / it possible, that non-engram portion is still FP4 of ~same size! Need to read tech report.
Actually this ought to run quite well with SSD streaming. The MoE expert sparsity seems to be similar to DSv4 Pro (hence exceptionally sparse) but with far fewer total and activated params. The added engram params can reside on disk as well (similar to Qwen Flash-Next), the additional load on storage performance will be quite negligible for typical scenarios.
By reducing per-session KV cache requirements even further compared to DSv4 Flash, this model likely opens up near-frontier model inference (in slow, unattended scenarios) even on low-end consumer hardware, as long as it has enough fast storage to host the model weights. This will be extremely exciting.
> It also includes additional 196B Engram memory which you can put on an SSD. I think
You can put Qwen 3.8 Flash Next engram on SSD, but prompt processing takes a good hit. On my mac studio, I get 300 pp and 33 tg with SSD offload, versus 550/40 with everything in RAM.
I will be very happy if 300 pp is achievable with this model though.
The engram stuff is great because RAM is often still cheaper (or at least expandable).
My company does currently look into buying some hardware as we handle confidential data and code.
Qwen 3.8 Flash is viable on two Nvidia 6000 96GB with a wood quant because you can put the 50GB Engram into RAM and the hit should be below 10% performance.
At least that is what I have seen so far.
Correct me if I'm wrong.
I am running that on a single 6000 96GB with 4-bit quants for both weights and PLE table. Needs just 32GB RAM and fits snugly into the 96GB VRAM with KV cache equalling ~300k context tokens. Not sure if I quantized the KV
You can warm cache regularly used engram/n-gram if you're willing to merge PRs into a personal branch and build it yourself. I was trying this with qwen 3.8 flash next and the n-gram to get it to fit on my very average gaming desktop (it worked)
While I instinctively dislike any "secretly spying" devices, phones could do that for a while now. I wouldn’t have much use for my watch doing it, but my nephew swears by taking notes on an ipad with an app that has a sort of a playback: essentially a play bar that recreates the note as it is written, with audio included. Being able to playback the lecture afterwards that way is a huge benefit over pure notes and I wish I had something like that when I was in school.
Friction makes all the difference. Google used to sweat over 100ms in page loading.
Running a voice memory all day long is one thing. Pressing a button that automatically summarizes every conversation you'll ever have is a step change.
There's a difference between "weird people could choose to do this creepy thing and violate my trust" and "not weird people could forget to turn off this slightly less creepy thing and still violate my trust" Both of these suck, but honestly I don't appreciate that Apple is making me feel uneasy talking around my friends with Apple Watches. Not because I don't trust my friends, but because now it takes extra work for them to not do something kinda creepy.
Jfc the tedium on this site. Sure, no one makes me feel anything, you're technically right buddy.
I don't appreciate Apple normalizing the already creepy ai note takers that exist and increasing the odds that someone will use them in casual or intimate conversations.
It would be shocking if it wasn’t trained on sessions. Have you read the ToS parts for both openai and anthropic that talk about it? It’s so obviously a weaselly way to say "no we do not train on your exact chats but we talked with legal and we think a cleanroom reimagining of your convo is probably fine and frankly where else are we going to get such a treasure trove of training data?"
There’s potentially trillions on the line, do you seriously expect those companies to adhere to laws and regulations any more than, say, uber?
The only unlikely part is the timeline - your sessions from a week ago probably haven’t made their way into the model. It’ll just take a while longer, and will be massaged just enough so that it isn’t really your exact session word for word so you can’t sure as easily.
at first I thought your post was a bit revolting with "have you read ToS?" bit, but in the end I completely agree and understand
I also don't get why it was downvoted, other than due to people not reading past the first sentence - although in the modern world's attention deficit that is understandable too
GLM 5.3 is probably the sweet spot open weights model if you want to go beyond deepseek flash or the new glm flash. I used it with pi and had a fairly good time, especially since it’s less touchy about cyber and whatnot than the US guys. It’s slightly behind Kimi in ability but it’s a lot easier to run it, I’d expect prices (and speed!) from third parties to be noticeably better.
Assuming you’re willing to drop a fat stack of cash on the upcoming Mac m5 ultra with 512 gb unified memory, you can even run it locally, quantized to 4 bit. Whether it’s even slightly reasonable, well, my wife would probably skin me alive but maybe yours is more understanding.
* Models are evolving quickly with high worldwide competition
* Hardware is evolving despite RAM shortages
Is investing a huge sum of money in equipment for local inference a wise use of money? Or are M5 Ultra and equivalently priced local inference hardware future-proof enough to be worth it relative to how the market is evolving? Maybe it’s all a question of what you’d spend otherwise on serverless or dedicated GPU spend…
It is absolutely not worth buying hardware to run models for purely (long term) cost reasons. For open weights models the economies of scale means the cloud beats local significantly and your payback time is like 10 years.
However there are other reasons (e.g. privacy) that might make it worth running locally for some people.
> I think the biggest reason is to own the stack so your model can't be changed out from under you,
The concern would be future regulations that prohibit you from buying a hosted version of the model. Even that could be bypassed with a VPN to another country but it's more work to go through the payments.
As long as there is demand for a model, it will be hosted by multiple providers.
I live in a place where using VPN is illegal and akin to "terrorism" because why would you want to hide what you are doing. Only bad guys hide. So if you use VPN, you are a bad guy.
“Out of the 15 individuals identified, five were minors who were counselled and advised in the presence of their guardians, with emphasis on awareness, lawful digital conduct, and the consequences of violating lawful orders,” he added.
What if the model is hopelessly obsolete, and thus no demand, but I want that specific model? Owning the weights and hardware is not just solving for one problem. It eliminates all the classes of problems that occur outside of your building, if you have a solar and battery setup.
Also, on a more practical basis, what if the way it's served is bad. Maybe I want my specific KV setup, or ultra low quant for entertaining garbage at 200 tk/s
> What if the model is hopelessly obsolete, and thus no demand, but I want that specific model?
You can still find a lot of old and completely outdated models on OpenRouter. The providers can scale serving of models up and down as demand arrives, so models don't generally disappear. They're just kept in the mix and the clouds will allocate hardware to it if someone is willing to pay.
In the odd case that it disappears completely, buying the hardware 2 years from now is probably going to be a better deal. That wasn't true if you selectively check the time period before hardware got expensive, but as new hardware comes out we're going to start seeing Strix Halo and old Apple hardware hit the market as people upgrade. It's already happening.
There is a certain personality type that cannot tolerate any uncertainty and must lock everything in right now against all future possibilities. If you fit that description then there's nothing anyone can say to discourage you from buying your own hardware, but for everyone else I do not recommend buying hardware to self-host LLMs just to save money. I self-host and run a lot of tokens through my setup (non-coding work) but I'm not really saving money.
> There is a certain personality type that cannot tolerate any uncertainty and must lock everything in right now against all future possibilities.
I thought HN banned personal attacks. I'm in this sentence and I don't like it. /s
I just buy the good apple hardware because it's good, and it also happens to run local models. It's not as good for the dollar, don't get me wrong, but I'm not going to develop iOS without a mac, that's even more questionable than buying a strix or whatever.
Also, this makes me wonder if, by using a bicycle generator, and a local model at sufficiently low power consumption, you could directly claim to have produced the text in a really physical way. "Yes, I generated the electrons that made that text work by my own efforts".
You do, there's like 20 providers for any model on openrouter. You can also just spin bedrock or gcp and download the weights for later if you're worried. It's never going to make cost sense when the token rate is so low with how expensive ram is
I think that's overly pessimistic. Here's [1] a video of somebody running it on a ~$6000 rig and getting around 14T/s for complex prompts (about double that for simpler prompts). Payback time is going to depend on your electric cost/consumption. In most domains cloud providers end up charging a significant premium rather than a offering a scale enabled discount, relative to local at retail costs. That will almost certainly end up being the case with LLMs as well, if it isn't already.
Furthermore we continue to follow the path that image gen neural networks took. In that domain hardware requirements reached a peak and then started sharply declining to where we are today where a plain old video card can rapidly generate images that took a supercomputer not that long ago. So it's reasonable to assume that performance of such a system could potentially even increase over time.
With roughly 2.7 million seconds per month, times 14 tokens per second, you are getting 38.5 million tokens a month at most.
That’s less than 164USD worth of GLM5.3 tokens on the inference market. So that 6000 USD rig will take 3 years to break even - and only if it runs continuously.
And this is being generous, as it’s not even taking quantisation into account.
I think the “killer app” is doing inference without sending the data to China or the US. At home it’s overkill but imagine you are an EU consultancy with a lot of client data to work on, or a company/institution with a lot of sensitive data, buying the hardware to make sure the data stays private is a huge benefit. So is that you “own” the model. Its capabilities, price or access don’t change at someone else’s whim.
Some of that is that EU providers need to up their game here.
Needing an EU native option is really the one and only reasonably objection I've heard against using LLMs from the cloud, the rest is tin-foil hat level unless you're actually intending to meddle with the inference or fine tuning or something beyond just querying.
I think if you steel-man what I'm saying, what you're saying falls apart. 14 tokens per second was rare. It only dropped that low in one scenario where he had it single shot an entire game (flappy bird clone) from scratch, with different assets, all self created, and so on. It ended up resulting in the LLM doing stuff like plotting out a some odd 100 item long to-do list, requerying it repeatedly, and so on. And it succeeded.
Also as the video mentions, the guy wasn't very familiar with what he was doing, and so there are almost certainly various optimizations on the config side he could work out, especially as he was using a 5 GPU system, which default configs are probably not well optimized for.
But I think we've rapidly moving along the same path as image gen stuff. Local generation has gone from purely theoretic, to requiring supercomputers to run relatively incapable models, to where we are today - where with a fairly basic high end setup, he's comfortably running a frontier level model. There's definitely an argument for going local that's only growing stronger by the day.
I agree it’s probably not representative token speed. But I do believe the overall observation holds: The monetary value of local inference is bound by the wall clock.
I agree that there are many other reasons than cost alone.
I'm actively uninspired to write high quality code when using Anthropic/OpenAI models given the high chance I'm a customer as well as used as dataset generation tool for them.
But currently cloud does beat costs of hardware ownership, particularly with ridiculously high RAM/GPU/SSD costs....again due to these same companies.
> It is absolutely not worth buying hardware to run models for purely (long term) cost reasons
This is especially true when it's trivial to have the LLM itself write you a script/tool that can rent a GPU node for you (via API calls to providers) and then download and set up an open weight model for you.
I mean, I think it depends. At home 3 of us we use AI for multiple reasons, from coding apps to asking general questions, and if we would have to pay equivalent subscriptions that would be ~1k a year on AI + submitting all your data to external services. I payed around ~8k on 2 DGX Sparks that, at the moment, serves perfectly fine as a ChatGPT/Claude replacement at home (DS4 Flash peaking at ~170 tokens per sec with 6 concurrent sequences), and even once the technology is obsolete for inference in a few years, I will still have 2 pretty powerful machines for whatever I need + some pretty fast NVME Storage. I don't think its a terribly bad idea.
I think the privacy argument that keeps coming up is overrepresented. Certainly ZDR is enough for an absolute majority of use cases? I see so much talk about local inference but I doubt most of it has privacy as a valid argument (not arguing it doesn't exist). It's fun to do things locally though. I've tried it as well but cloud is just faster and cheaper.
These companies have displayed zero respect for everyone's intellectual property getting these models trained.
I think not giving them your complete trust is reasonable! I'm not saying zero trust, and ZDR is fine for most things but I understand the people who don't want to stream their whole codebase out token by token.
Then use other providers hosting open models. Companies and individuals already put their whole code base on the cloud. I'm genuinely interested in privacy-oriented use cases where ZDR is not enough.
ZDR is built on trust. Given that end-to-end encryption fundamentally doesn't work with LLMs, as they need the content to be unencrypted to operate on it[1], you have no way to prove that once your plaintext data is on somebody else's server they aren't doing whatever the hell they please with it. All you have to rely on is their pinky promise that they won't do anything with it. Trust is a valid option, much of our society runs on trust, but you can eliminate the need for trust whatsoever by running on your own hardware.
[1] Yes, I'm aware of experiments to operate on encrypted prompts, but these are only research attempts, not something that could actually be used with frontier models in production.
I'm not that worried about the codebase itself. I'm worried about the fact coding agents poke around the terminal and system so much that there is almost a certainty that some of your other personal data ends up in the context somewhere which is getting logged in to a training dataset by random hosting providers.
Privacy isn’t only, I don’t want anyone to have access to my data. It could also be, I don’t want anyone to know my use case because it’s niche and highly profitable.
I have a Strix Halo and dual 32GB GPUs in my desktop, that sit idle right now, because the electricity to run them and to cool them in 110F weather Texas is currently experiencing pretty much nulls any savings I might see over getting better models from cloud providers. While I mostly use Claude or Codex with subscriptions for agentic work, for API use DeepSeek has usually been my go to, but now I guess it's GLM 5.3 or the Flash version. And, for security work that Anthropic or OpenAI models are likely to refuse, I've been using Kimi K3 (also via subscription, though their subscription is extremely stingy), but I guess GLM is now the one for that, too.
Anyway, yeah, even at the prices I spent on my local AI stuff (I bought before RAMpocalypse really kicked into gear, so I bought old server GPUs for about $350 each and the Strix Halo for a little over $2k) it was never going to pay for itself; I just like to tinker. But, I can't imagine spending today's prices for hardware for local AI.
When the memory shortage ends, I'll be down to the Apple Store (or, more likely, clicking refresh on the Apple outlet every few days). But, until then, there continues to be a glut of cheap and free models in the cloud that are better than anything I can run locally and they're faster, too.
Yeah, I guess, but it feels like there isn't really an opportunity for anyone to do that, given how competitive the market is. If Anthropic decides to demand API rates for everything (which would make my $100/month turn into a few thousand, I guess), I won't be seriously inconvenienced by switching to GPT. And, if both of the major American providers do a pricing collusion and GPT also becomes thousands of dollars a month to use, I can choose between Kimi K3 and GLM and so on. I'd rather use Opus 5 and Fable, but I'm not going to be seriously put out if I can't. We've got three or four open models to choose from that are as good as or better than Opus 4.8, which is Good Enough, and the competition isn't slowing down. We're seeing more new competitive models more frequently than even three months ago.
So, even though there are more models to run locally that can be useful for the stuff I do, it makes less sense now to do so than it did when I got it. There are more extremely cheap options, now, and it seems likely to continue to get cheaper and better and faster, while my local hardware will always be slow and hot and only gets better via software (which has roughly doubled the speed it can run stuff since I got it, but it seems like there's less room for improvement in software now, and even at twice the speed, it still kinda sucks to use local models interactively especially on the Strix Halo).
If they "should" cost 4k in the sense of marginal cost, then you will be spending more running the same at home, because your home hardware will always be less efficient.
There is a big difference in the cost of a 5-nines up time system in a heavily space constrained environment compared to a home hobby white box used for some coding. The GPUs alone cost 10x for the data center versions compared to the gaming versions even with similar specs.
The cost of online services is also largely a result of the cost of training (though hard to say exactly what that number is). Assuming you are using open weight models at home, you aren't paying for the training - someone else is.
> The cost of online services is also largely a result of the cost of training
OpenRouter prices are somewhat simmilar to Antrhopic/OpenAI API prices. So I conclude that the hardware plus operating margin alone can genuinely produce prices way above what you'd pay if you had a subscription.
Of course the primary unkown factor is average token use per subscription. Without that it's all wild speculation.
I wish Texas would write up a regulation allowing 'balcony solar' as I could easily generate 1000-2000w of solar in my small back yard to take a bite out the sizeable cooling bill I have.
Seems like it's easier to ask forgiveness than permission. And, I wouldn't bet on this legislature ever doing anything that would disempower fossil energy or reduce their profits, even a little bit.
There are good reasons you aren't allowed to plug random power generators into the grid. You might be allowed to have your own ones not connected to the grid. Remember that graphics cards run off poorly regulated 12V, although you'd want to regulate it anyway because they're expensive to replace if I'm wrong.
Balcony solar by definition plugs into the grid. It's not just "solar panels that happen to be on a balcony", it's a certain type of self-contained unit that plugs into a regular power outlet.
Nope, because of laws written by energy company lobbyists for the "safety" of consumers, so if you're mixing grid and solar on the same site (even if completely isolated circuits) you have to get all kinds of certifications and permits for inverters and breakers installed by certified technicians that make it cost prohibitive.
Also, many HOAs and local governments place ridiculous restrictions of their own right
I posted more on the other branch of this thread. But cost prohibitive is about all I can find about this and not illegal. Which for some people might be the same result of course.
As I understand it. They banned solar inverters as well as severely tariffing the panels. In some places it's also illegal not to get your electricity from the grid.
So far I don’t regret buying an M1 Max device with 32Gb of RAM. The models available for it keep getting better (running just about okay for interactive use) and 400 GB/s of bandwidth is still considered a lot.
The models are currently improving much faster than the hardware and this doesn’t seem to have plateaued yet.
Try 3.8 27B in MTPLX; I get about 30 tok/s with the same hardware as you. (Although it does use around 90-95W of power, compared to the ~60W that 3.6 35B-A3B uses to generate 55 tok/s. That’s about 3 J/tok instead of 1.)
NGL: I don’t really have a good way to find out right now. It also doesn’t matter that much because the way the models use the tokes varies a lot. Qwen 3.8 is known for overthinking while Muse Glimmer may be a little slower per token, but it uses them very efficiently, caveman style.
Generation speed isn’t the bottleneck anyway, at least on pre M4/M5 devices (the newer chips got significant processing acceleration). It’s prompt processing time. OpenCode’s system prompt can take up to 3 minutes to process, which is why good prompt caching is essential.
For that I use omlx, which can persist the KV cache to disk, chunked so you can reuse parts. This helps with the usability a lot, when an agentic session is warm it runs pretty smoothly. New requests can take a couple seconds (sometimes many, which must be fixable somehow).
So: It’s not fast, but I also don’t find it awfully slow. My use is typically semi-interactive, for fully interactive use you have to wait a bit, but it’s possible. I personally am still regularly amazed that something even close to this is possible on completely local hardware.
That's basically the question I'm trying to answer.
If you're paying Anthropic or OpenAI to use their models, harness, governance, etc., I could see the local inference potentially coming out ahead. They're already starting to ratchet down what your money gets you on their platforms, and that can be expected to continue as the leaders of those companies continue to seek the road to the El Dorado that is being a trillionaire.*
If you're looking to get into the guts of AI development instead of having it handed to you by a provider, that's where it gets murky. I'm wanting to write some sort of agent that does things and get into making outputs consistent in the like, and I'm not sure whether to host something on GCP or buy an M5 Mac.
*Note: El Dorado is a mythical city and many people died trying to find it.
With competition we kind of have guarantee up to what providers can do, they don't have that much control, the most radical thing they can do is to go bankrupt.
Part of it is knowing that whatever sort of enshittification the cloud providers do, my local programming environment won’t ever be less effective than it is today locally. It’s the same reason my entire development stack from editor to compiler is open source. I don’t need to modify it today, but I always must retain the option to do so later.
There are several things I do in my life that only pay off in the event of a big disaster, like an extended internet outage, civil unrest, supply chain disruption, war, etc.
I like to be able to do the things I do even if offline for weeks.
I spent a lot of money for more flash in my iPad Pro so I can keep all of offline wikipedia and OSM in it, for example, along with tons of books. It’s sort of like being a digital prepper. (Being a prepper is a spectrum, from anyone who keeps food in their pantry to people building bunkers under their house - how much you invest is a personal prudence and threat modeling decision.)
Also, privacy. And when I got the Mac Studio the 512GB was only $15k, which is dirt cheap for that much VRAM.
> Part of it is knowing that whatever sort of enshittification the cloud providers do, my local programming environment won’t ever be less effective than it is today locally.
Is that true though? Many of the core LLMs need to be retrained as languages evolve to incorporate changes (language specifics, compilers, tooling, etc.). To some degree this can be handled via context injection in a variety do forms (agents looking up documentation and so on) but inevitably it’s not stationary in time, just as your OSS stack (probably) isn’t (depending on the languages, technologies, and use cases).
So your hardware is to some degree dependent on the good merit of groups like Z or Alibaba or whomever pushing out updated open weight models that dumped loads of capital into to train. You can keep using the existing models but at some point I suspect they’ll start to have more friction due to dated specs in language and so on. Again there are tuning and ways of layering this information on, and in theory you can even do some training on your own but I don’t think it’s as stationary as being portrayed here.
Those updated open weight models may not always be there (updated on new data). The usability of them is probably fairly long to be fair, but I suspect you’re going to see explosion in everything from libraries to languages etc due to LLMs so even the rate of change across your OSS stack may cause these models to be dated quite quickly, at least in the core model which will require layering fixes.
To be clear I’m on the fence thinking about much of the same issues and as close as I am to pulling the trigger, I keep thinking of very valid counter arguments as to why it’s me just wanting this thing I own. Which may be enough.
With every newly released open weight model, the clock on the issues you describe is reset. I can see a marketplace arising for paid updates to common lines of open weight models, which will incentivize those with the hardware to train to fix the problem for those who only have the hardware for inference.
I would say when this comes to pass, we are already 5 years along?
> Part of it is knowing that whatever sort of enshittification the cloud providers do, my local programming environment won’t ever be less effective than it is today locally.
I think this is quite understated. It basically is freedom from a growingly antagonistic relationship between you and some remotely hosted API managed by faceless corporates at the whims of their board, shareholders and governments.. It really is such a mental burden to need to constantly manage this relationship (watermarks, silent downgrades, random false refusals, downtimes, model sunsets, changing ToS's, fucking ads). These companies will need to squeeze you for every cent that they can before open-weight models are simply good enough for the valuable tasks we can throw at them.
To have your own hardware is to no longer have this mental burden.
Only reason to spend a bunch of money on hardware to run LLMs locally is if it's a hobby to you to an extent that even renting the GPUs temporarily won't satisfy you.
Or if you need stuff that APIs don't / can't provide. Or for future proofing your workflows. Running things locally gets you "the same thing" in perpetuity, while APIs might change, models can be deprecated and features can be removed.
Cybersec is also hit and miss, depending on what provider you choose, verification systems and all that jazz. Also, running locally allows you 100% data privacy, in any situation and for whatever usecase you might have. ~100k for hardware for a small team of devs to code locally is not that expensive in the grand scheme of things.
Lastly, local models allow for training / finetuning on your own data and processes. $/tok is not everything for everyone. Sometimes you can take a hit on value / speed if you get something else that matters for you.
Or Qwen 3.8 27B, AA score 52 (which is utterly insane given the size of this model); I have been testing Qwen 3.8 27B since a week now, as an intensive GLM-5.2 and Opus 5 user - I can say that I just can't believe my eyes i.r.t. to how good this model is.
I have just built an Epyc with 512gb DDR4 3200 RAM for a "reasonable" price and I'm hoping to have a setup with GLM as the architect and Qwen 27b/Next Flash as the implementer. This is 1/5 of the price of the Mac, but also probably 1/5 of the speed lol.
I’ll be very curious what you get with DDR4. I also almost went that way. I have an Epyc DDR 5 rig and the best I see is 10 tok/s. Caveat being that’s at Q8 and a 4090 doing pre fill so it could be pushed up.
The surprising thing for me is how much work you will need to cool the banks if you’re near your memory ceiling. My memory starts soft throttling at about 74C (dies may be hotter, that’s the bank temp) and will turn down speed to try to stay below 80.
Happy to send my llama.cpp config settings if you want it.
I am getting 10t/s on unsloth's Q3kxl with 2x3090s@250w. It's enough for me for now. I will probably upgrade the GPUs down the line. DDR5 would have made the price of the machine double and I just wasn't prepared to pay that much.
I find 10 to be very usable. It’s not (that) interactive but it chews through tasks. I let Kimi churn away at 4 overnight and it gives good results that are ready for me in the morning.
Typically computers with these larger memory amounts have fans that scream like a banshee trying to move impossible amounts of air over the memory and CPU. Getting something both cool and quite can be a bit difficult.
Yes, I thought when I was starting that 1u and 2u form factors were to save space. Maybe they are, but they also have the advantage of moving air front to back very effectively through and over the components. Though there still must be some need because I see even those boxes have optional manufacturer built memory shrouds to try to force airflow between the DIMMs.
I had some 120x38mm fans from another server box that I pulled out because they were too loud and I didn't need the static pressure they were giving. They went in here. That 13mm (and the extra 1k rpm) moves so much more air.
I built a dual epyc server with 64 cores and 1 TB of DDR4. Draws around 800W or so under load. I used off the shelf liquid cooling. It is audible but not noisy.
The trick is to turn on the cooler's RGB in your 6000€ server to get a free speed boost. I am not liable for sysadmin's heart attack upon reading this.
Honestly I suspect neither of them will be performing terribly well but with DDR4 3200 RAM I wonder if you'll be counting tokens per second or seconds per token. I mean, you do at least get a lot of memory channels at least, compared to consumer PCs. I am curious to hear what performance you get, I feel there is not enough information out there on what different setups manage to eek out.
The fastest I was able to get my Threadripper 3960X + 2x 3090s + 256GB DDR4-3200 to run a 2-bit quant of GLM-5.2 was 8 TPS. I would expect to be in seconds-per-token territory for a pure-CPU 4-bit quant.
One thing I'd like to try is MoE offloading: I have 2x32 GiB of VRAM and 128 GiB of DDR5 running at 4800 MT/s (only 2 channels though). I've seen people post difficult to believe MoE offloading results albeit a decently long time ago with older models. Maybe there is a quant that would fit with MoE offloading?
That said, I am guessing my problem is not enough RAM - but this poor consumer platform struggles to do memory training with 128 GiB as it is.
Now I surely regret not having gotten Threadripper and 256 GiB of RAM in the before-times.
My measurement was with MoE offloading, but there's only so much you can keep on-GPU with a 200GB quant and 48GB of VRAM. It's hard to overcome the CPU/RAM bottleneck.
For what it's worth, all of my hardware was used; I think, all-in, I'm probably at around 3k-4k USD? Not cheap, but also not the worst for something relatively versatile.
With a 4-bit quant of GLM-5.2, I can get about 0.8-1.1 tok/s on an underclocked dual Xeon E5-2698 v4 with 512GiB of DDR4-2400. I think it was specifically a Q4_K_M quant. Of course, the time-to-first-token is absolutely atrocious.
Which is completely insane for a ten year old configuration.
What model are you interested in? DS Flash 0731@Q4KXL I'm about 25-30tps. Same as the new Qwen3.8 Flash Next. The new GLM 5.3Q3KXL at 10tps. I've got 2x3090s which I didn't mention in the original message.
Well if you did get the m5 ultra could you obliterate the guardrails and then your wife can ask it pertinent but unsafe questions about how to punish you. Seems doable.
Though I appreciate how good it is at "solid" grunt work and at that price (in fact I am paying the grandfathered subscription price; mostly).
I am planning to let go for my Claude AI subscription which I now use only for "planning" and maybe use that via Open Router as PAYG (at to try how it ends up). But god glm is bad at "talking" and "responding" anything prose. Not only quality but it's almost impossible to tune it and make it let go of its habits and biases and enthusiasms which often result in too many too and fro.
So I sometimes wonder at what point that starts becoming the cost and mental hassle. Maybe it's not there for me yet.
One could also run it locally on a used dual xeon (or amd-equivalent) server with 512GB RAM, albeit slower, if you have a useful workflow for it that's like "take this day's efforts and run it through various analysis agents", combined with giving it one-shot tasks/modules to build overnight. You would want a place like a garage or basement to put the server because it'll be loud.
Memory throughput performance is better than many equivalently priced (talking about the cheaper end of the used/refurb market here) single socket systems because each CPU is talking to its own banks of RAM.
Using an example of llama-server on a 256GB (each CPU is attached to 128GB) system running qwen 3.8-flash-next Q8, which takes up about 188GB with the model and context, it's not that the inference engine 'knows' about NUMA, but if you use --load-mode mlock to load the whole thing into memory at once at the time llama-server launches, the kernel will distribute it relatively evenly across the memory connected to both CPUs.
I was going to try out pi and I set it up with some of the recent big open weights models, but then I realized that if I just use pi the way I use claude code, it doesn't have an auto mode so it's basically just --dangerously-skip-permissions and you're supposed to sandbox everything yourself. What's your sandboxing setup like?
”less touchy” is an understatement, it happily complies with running exploits, reverse engineering and decompiling. Asking Claude to do that will give you an error and make you feel like you’re about to get banned.
There's certainly a mental difference between a computer you can use as much as you want for a fixed upfront cost vs a rented server you are being billed by the hour for.
But the cost difference between api and self hosted is so incredibly large now it's almost impossible to ignore the fact self hosting is a terrible deal. I'm waiting things out until the dust settles on what the required specs are and consumer hardware gets cheaper/more capable.
and even if you were making such a salary, the quesiton of if the investment on hardware to run llm's locally is still a big if, its OK if you buy the HW cause you'll use it and you get the extra capability as a nice extra, but doesnt make sense to spend so much when you could just get 200$ subs with almost infinite SOTA tokens a month etc (if you dont need the local/privacy aspects of it)
App vs website is purely taste. Some people will never install an app but will use your website. Some people love apps and will never use your website. Some people prefer apps to PWA when the app does thing, because what if it changes and enshittifies or just vanishes - an app doesn’t need to be updated. Some people prefer PWAs to apps because why would I bother with the endless updates just fix it on the website, in fact I don’t even want your shitty app at all why in Lord’s name is this an app?!
Sometimes… well sometimes there’s no dev resources available to make an app work in a way you’d expect. Google is just a tiny startup, be reasonable in your expectations.
Point is: if you actually want to reach a large audience — you want both. Also, don’t expect good experience if you use Google’s products.
Anyway. Routers are such juicy targets, especially since there's less eyes on them than on laptops/desktops, that it's hard to believe there are any which aren't backdoored, be it by China, US, US' middle eastern sidekick, whatever. My question would be: how limited are they to software, like in the article?
You can flash with OpenWRT and Fritz, but to my knowledge it doesn't replace the boot parts. A suspicious bootloader in SPI flash probably isn't difficult to produce, even if it would likely be discovered... eventually. A separate modem / radio processor would be an even bigger nightmare, esp if it's an 5g router - you essentially add a second computer to your router, with its own components and network access, and I think you can't even play around LTE/5g networking as freely as with wifi without bending some laws.
Theoretically you could watch the router from, well, outside the router. But as the article says, it can be difficult to say what's an actual connection and what's uploading your network's data to an uninvited third party. I'm also wholly ignorant what implications it has for things like Wireguard and Tailscale. I assume you're completely pwned if the Wireguard runs on the router, but I don't think it does anything if it runs on your laptop or whatever? Then again, there's a lot more devices in peoples' homes nowadays, and almost all the IoT has pretty much zero security.
I suppose AI will help with some of those things. Tracking every connection made is very tedious for a human, but throw some compute at it and you'll get a much better insight, especially since you'll probably know what to expect out of your machines.
I guess using a mini PC as a router-slash-proxmox-hub would be a safer choice? There's probably many fingers in this pie as well, but my gut feeling is that it's a lot safer than some cheapo box provided by your ISP at the cheapest price they could source them.
Probably safest to take a layered approach. Something with long update timeline and verifiable source. And then additionally use encrypted communication on your devices behind that router (DNS as well)
> 320B total parameters and just 18B active parameters
This is pretty hefty for a "flash" model, even a 256 GB setup is insufficient at q4 - and q4 is already the worst-but-still-acceptable quant in my experience. The benchmarks look great, especially since GLM tends to be more honest than the average Chinese lab, but you’ll need to splurge to run it at home.
@edit: so many releases that I forgot to math. This fits just fine in q4, realistically the minimal hardware would be 192gb - so blazing fast on double rtx 6000 pro and usable on 256gb unified memory. You could even go with 5bit quant on 256gb.
Speaking as someone who isn't really well versed in this, does 18B active parameters mean that you could potentially hold only the 18B parameters in RAM and stream the rest from a fast NVMe SSD for acceptable performance similar to how Colibri works?
Normal MoE is switch-weights-per-token so you would se substantial slowdowns that way. Apple did a More that switches weights per prompt (instruction-following pruning, https://arxiv.org/abs/2501.02086) but you have to design the model that way which I don't think the have.
Sorta, but you're off by one layer. You can store the 18B in VRAM and stream the rest from RAM. There's still a performance hit relative to storing it all in VRAM, but it's tolerable.
Generally, for local consumer use, these large MOE models are best for unified RAM systems like DGX Spark or Mac Studio.
People are annoyed at apple’s recent lack of polish in software, but the increasing monetization and push into services are a bigger issue imo. You pay the apple premium for hardware and the hardware is indeed great, but you’re increasingly treated like a piggy bank. Line must go up.
In regard to Maps: if all you want is a gps, there’s many apps that use OpenStreetMap data. I use CoMaps myself. You download the map(s) you want and it just works, no forced updates or ads or anything in the way of upselling.
Strangely enough I often use Google Maps on my iPhone because I'm looking for businesses that have no presence on Apple Maps.
I don't see the opportunity for companies to present themselves in Apple's app as a net negative for users.
The margins in the services category are much higher (76.7%) than hardware (38.7%) [1] and much easier to grow in an already saturated smartphone market. So yeah, time to throw the customers under the bus.
when I switched from Windows to Apple the amount of ads was the same. The only thing that changed was that they were Apple's ads. Apple hardware is great, but it's being severely botched by it's software. Seems like software is hard, no matter how much money you have.
>And what happens when the price goes up by an extra zero on the end?
>And then another one?
Well so far the price trends towards going down by the zero at the end. Inference isn’t expensive, there’s a whole laundry list of providers to choose from. It’s training (for everybody) and salaries (for US labs) that are expensive.
>and you have to switch to using local models?
Many companies are interested in that, if for no other reason than to avoid uploading secret sauce to we-totally-dont-train-on-your-chats-pinky-promise API endpoints. A mid sized company can cough up $500k for dxg b200 which will be useful for years and years and the only ongoing cost is electricity. Put glm 5.3 on it and it’ll zip along for a number of employees at the same time.
>And then, suddenly, those local models cost an absolute bomb to license, because the companies that bought up the wreckage of the original model creators are trying to recoup their spend?
How are they recouping anything on a mit licensed model released into the wild? There are plenty of models released by "who?" companies and the cost to download those is zero.
>Meanwhile the company is being hit by nuisance litigation from people whose code the bots plagiarised
Any day now
The doom scenario makes no sense in the race to the bottom that’s happening now. China being the other big entity that’s invested in AI makes all the what-ifs about litigation and plagiarism moot; they will gladly offer everybody inference for dirt cheap, if for no other reason that they’d like it if you gifted them your IP that they will then use to outcompete you with their manufacturing prowess.
The genie’s out of the bag. AI is never going away, and we all need to adapt whether we like it or not.
reply