The nice thing is that current local modes don't go away. The weights are out. Qwen 27b and deepseek are already very useful. Hardware will keep getting stronger so after the current memory crisis you can run them more cheaply.
Of course the proprietary models will continue to improve and we have no guarantee that there will be new open models. But for coding, the local models are already good enough for a vast amount of use cases, if you are an experienced developer yourself and can guide it.
On current GPU hardware/firmware... The cynic in me imagines in the future GPU vendors wanting to implement only loading signed weights to your card, aligned with your "subscription level", for your "safety".
But the GPU vendors won't really have an incentive to do this - artificially restricting their customer base? Unless they and the hyperscalers all collude to cartel the market of course!
This is why it's so important to have more than one hardware vendor. The bigger the ecosystem, the harder it is to cartel. I cant believe it, but i am pinning my hopes on china to break through with commoditized chips and hardware.
> But the GPU vendors won't really have an incentive to do this
Unless the GPU vendors start selling end products? It's much more lucrative to sell locked down end-user devices after all. Shareholders will demand it.
I was just implying that the (free, unregulated) market does not always give the consumer what they really want. With two examples where demand does not lead to consumers "winning".
The average/median consumer may get what they want. However, unfortunately in practice it means that lots of people will not get what they want. It's like the tyranny of the majority that some people believe is a part of democracy.
People came up with the word "enshittification" for a reason ... and the word is quite popular on HN, so make of that what you want.
I love the Lotus Elise and desperately wish there was a similar affordable, super light weight; no frills, sports car made today.
Here I thought I was just a niche market that’s not profitable for manufacturers to serve, but apparently I’m a victim of tyranny and didn’t even know it!
That's a fair comment. There is lots of speculation around NVIDIA is shifting all GB202 production to RTX Pro GPUs, away from the consumer 5090s. So maybe less about restricting user base, and more about focusing on where to maximize returns.
I worry they get so much money from hyperscalers we just get scraps. Consider that I'm running 6 and 8 year old GPUs for my local inference. If I wanted to replace my hardware with latest stuff (nvidia dgx, gorgon halo, Intel b70 or amd r9700 pro) it would cost me 4x the price and I'd cut the memory bandwidth in every single case. In case of dgx/strix were talking less that a quarter of memory bandwidth and a third of compute of a single rtx3090...
Those products are a sick joke played on us. About the only thing they have that is better than tjose ancient gpus is energy efficiency.
And what good are the weights when all the rtx3090s fail (they were pretty crappy to begin with) and every ddr3 Xeon system fails too?
I'm fully prepared to be running on my hardware that is 6 and 8 years old for next decade. Last 3 months saw rtx3090 going up by 20% each month.
There is already a big market for open model hosting and the accompanying demand for people and DCs. If you know how to deploy models — yes even open ones — and scale inference you are a hot engineer in demand.
You don't need bigger or better models. Current models are more than adequate for any task.
The key thing is tool use.
The model is fixed but tool use introduces new information into the system, so once you have a model that is good enough to effectively use the full suite of tools available (debug, test, profile, etc) and write code to effectively combine them and synthesize their results for a particular task, there is nothing that it can't solve.
New models are continually ingesting new information from usage, and embedding the results from people running this process back into the base model, which makes it more efficient, gets more things right the first time, etc, but there is no step gap in functional outcomes left.
Computing and programming are a limited domain with a small fixed set of constructs that can be applied in an infinite number of combinations. Nothing ever closes the gap between the ability of the model to automate the assembly of these constructs and the necessity of expert human direction to pick what system to build out of the infinite range of possible systems.
Frontier models can keep producing very expensive one-shot demos to pitch to investors but they are never going to replace engineers.
> You don't need bigger or better models. Current models are more than adequate for any task.
At the very least we need _current_ models, though. The weights have all sorts of information baked into them, like "how do Python unit test drivers work" or "what does idiomatic C# look like," that will just naturally go stale over time as the frameworks improve. In five years, Qwen 3.8 will still be targeting the Python of 2026.
So target stable languages that don't change, TCC with win32 is a fantastic target. The apps work great with wine and you have opengl, controller and audio support. I have a wrapper for winsock 1.0 I need to package up and release, works great for small network apps.
An interesting question is just how much high-level programming language development we'll see going forward. What is the incentive to create new languages? They're all just object code to an LLM, so when starting a new project, you might as well go with an old-school language like C++ or Java or Python or JavaScript that is well-represented in the training distribution.
Why do we need Rust, for instance, when LLMs know how to use C++ without shooting themselves in the foot? If that's not entirely true today, it certainly will be tomorrow.
See also the lack of new instruction-set architectures in the CPU business. Why do we need more, in the absence of fundamental and revolutionary hardware innovations? Even RISC-V is just a rehash of MIPS.
I think someone will come up with a language which is nicer for "vibe coding".
I.e. I think you can bake certain properties into the language such that various kinds of static checks can prevent certain types of errors. This should make the language nicer for coding without reading any of the code.
You think LLMs don't make mistakes? My sweet summer child! I want all the guardrails with LLM code. Reviews are fatiguing. If I can remove "memory safe" from my list of worries, so much the better.
You know the LLMs that are finding all these CVEs lately are finding them because they are better than the people who wrote the original code, and better than every one of the people who ever laid eyes on it.
An LLM doesn't need a memory-safe language. It needs one it knows like the back of its metal hand.
Not really - fine tuning can be done on much lower tier hardware, with much smaller datasets. With a model like Qwen 3.8 27B as the core, a small fleet of LoRA adapters for whatever specific use cases you need would be incredibly competitive with the current frontier. Sure, future frontier models will likely continue to improve, but will people need them for everyday usage? I'd wager not. Unfortunately, my wager is dwarfed by OpenAI's and Anthropic's.
This is mostly a memory bandwidth issue. Qwen 3.8 27b and the related Bonsai 2 27B run on my M5 Max laptop with 36GB of ram. It’s a little slow due to my specific model having lower memory bandwidth than higher end models with more ram and cores. MoE models perform better here and expert streaming engines can allow larger than memory models to run.
Multiple teams and efforts are now making Qwen 3.8 27b work much faster on Mac and CUDA and the inference speed is doubling almost every week.
There's now ongoing competition about who will make it work faster with exactly the same byte output at https://mlx.fast . TensorFold is the project to follow.
They aren't "Just as accurate". What you see with these quants is the providers only ever show performance on already saturated benchmarks. You'll see labs put out 1-bit quants Deepseek v4.1 claiming only 1% degradation because they're only showing the benchmarks which tell that story. Then you look at the performance on those same benchmarks for significantly weaker models like Qwen3.8-27B even and it's virtually the same because that benchmark is saturated.
I like these quants as much as the next guy when they're all I can run but I have both R9700 with 32GB VRAM and a 5060 Ti with only 16GB VRAM and so I run both Q6 Unlsoth quants and ISTA IQ3_XXS quants and yeah there is an absolute world of difference in realworld performance.
IMHO I think it's pretty scummy behavior from the quant providers to frame these quants this way. Just be honest about where the degradation is really occurring on the non-saturated benchmarks.
This is a really good point that goes the other way as well. Some use cases are great for smaller and cheaper models, but it's not obviously when that's the case so I can bank the savings.
All I know is that Strata and these models are the most I have ever gotten out of my machine. I get 70 t/s on a standard gaming rig 9070 XT / 5900X / 64 GB DDR4 @3400.
Yeah, they're surprisingly usable eventually. What I found running the Qwen3.8-Flash-Next IQ3_XXS on CUDA kernel development was, sure it generated faster t/s than 27B dense, but it generated so many more bugs, and when the quantization is lower the reasoning traces get longer. It was more or less always able to figure out what bugs it had made and fix them, but it took quite a lot of time to do. I switched back to Q6 on the dense model and that just went away for the most part. I recently upgraded to 128GB system RAM so that I can up to Q5 on Flash Next. Really excited for Qwen4-Flash when it comes out. Gonna be a real beast.
I tried running one of the Qwen 3.8 models locally. Asked it to code a simple thing. It pontificated on its place in the universe and then asked me what I want it to actually do!
Claude would have written, documented, and unit tested the thing on the first shot…
I want to run local models, but if this is the current state, it’s hard to mentally justify the hardware expense to only get this…
4 bit has always been accepted as the quality floor that retains most of the models capability. Going beyond that brings problems. It doesn't make models totally unusable, but my contention is the quant providers are intentionally not showing an accurate picture, and then everyone is glazing 3-bit quants. They are what they are, but it'd be nice if we had a more accurate shared picture of it rather than just hype and vaguery.
Apparently the ud-q4_k_xl version of qwen 3.8 27b is quite better than the q3 and not much larger. At least per r/localllama. (I'm also trying to find a good local model).
Thanks, I haven't used the gemma models in a while. Last I remembered, their small models weren't bad but obviously showed signs of being "small". (Funnily irl I use Gemini 3.1 pro more than all other models combined).
Its possible (though very hard to do well, and needs a fair bit of compute) to improve the local models we have, too. Continued training is possible as time goes on.
Even if Deepseek v4.1 is the last open source LLM I have access to, I would be perfectly fine with that. My secret is that I have this relatively unknown technique for solving problems where I use this quantum engine that developed from just atoms over 4.54 billion years into the most sophisticated machine in the known universe. It's a frontier model that every human has access too, actually.
So when I really want to solve a problem, I just think on it. And when I need something to iterate through the things I thought of quickly, I use deepseek.
Yep, fully satisfied with DS 4.1 Flash or GLM 5.3 Flash. I prefer using one chat at a time, actually thinking through things with it, as opposed to the YOLO run all night multi-agent thing. I mean I’m still okay with subagents if my main thread needs to explore a few options and then report back, but all on the same idea.
You might enjoy Cecli, the successor to Aider. It's made for the kind of workflow you are describing (full hands off yolo and recursive sub-agents also an option, but not the default)
For this benchmark i 've been using OpenRouter. If you have enough RAM or you can ssh a cloud shell this intro https://memdoor.ai/docs/open-weight can work too. I have reached with the ssh door a Google Cloud Box where the 27B model ran with an L4 g2-standard-16. A inference engine server is setup easily with memdoor (https://github.com/guregodevo/janis) on any CPU or GPU box. Ideally i think we need a quorum of SSH boxes like this, distributed that you can ping with a mention such as @box:coder. It's WIP . Any suggestion for running a team of workers in remote machine is welcome. I'd like to find a cheap solution.
> Even if Deepseek v4.1 is the last open source LLM I have access to, I would be perfectly fine with that.
Your quantum engine aside, current open-source LLMs will not age well. They have a knowledge cutoff. If you ask them five years from now about something that didn't exist when they were trained, chances are they won't be able to provide a useful answer. They may still be useful for many applications, but their value will eventually deteriorate. Having continuous access to up-to-date open-weight LLMs is therefore about much more than simply having access to "better" models.
A distributed, open scraper specifically for LLMs would be interesting . Filtering for spam and useless data would be a challenge at scale, but would be cool
The thing here for me is not that I’m unable to solve problems with my brain, it’s that I value the automation in labor provided by coding agents. I do not want future advancements to be locked behind proprietary doors.
I agree with that but the cat is out of the bag now. Unless you believe in exponential improvement (which I don’t), the open models might be at a resource disadvantage, but that doesn’t mean that they can’t catch up or surpass the closed ones. Throwing money at a problem doesn’t guarantee success.
I mainly just want to point out the value of human ingenuity coupled with a model like DS4.1 is technically enough to get most people to where they want to go, and that includes smarter open models.
I do believe that the exponential improvement will continue at least for another year and likely farther than we expect, such that today’s open models, while being very impressive by today’s standards, will soon show their age as they are eclipsed by newer models in 6 and 12 months.
You say that now, but you don't know what the future models will be capable of. If models get ridiculously good, you might not be perfectly fine with that.
It can run on one strix halo machine, which is what I have. For everything else, there’s a vibrant community brute forcing more efficient engines and quantizations. There’s a minimum bar right now, so it depends on your hardware.
I bought my box for $3k last year because I hadn't updated my PC since the first threadripper 1950x, and the Strix Halo is the only x86 with unified memory that's optimized for the kind of CPU intensive stuff I do, and they finally have proper AVX512 too. It was the perfect upgrade for me (except for lack of PCIe 5 & ECC), which coincidentally happened to be perfectly positioned for local inferrence.
and I liked it so much (and saw that the RAM drought was going to last for at least a few years) so I bought a second one to use as a general server for myself and my family.
I’m all for local models and getting away from Anthropic/OAI but Qwen 27B might as well be a horse drawn carriage compared to a 2026 Porsche. Like, I guess it’s useful if you had nothing better but acting like it’s comparable to anything the frontier models have is just disingenuous.
disagree, it’s absolutely comparable. but an m5 max is where it shines. and there are 50 flags to finetune with llama server before you get it right. i do prefer llama server over mlx currently on mac as well.
What, I run it on 2 H100`s with vLLM and this is where it really shines. It's about half as good as Sonnet 5.5. On a small m5 max I can only assume trouble.
m5 max with UD-Q4_K_XL and 260k context is what i'm working with. and using compaction push it past 260k context just fine. i found when limiting the context that was one of the bigger debilitators. although, your setup sounds obviously better.
I have all three and you're not going to convince me of that.
Side-by-side Qwen 3.8 27B is ~6x as fast as DS4 and GLM 5.3 and faster than that if you get the speculative engine set up properly. That means you can iterate very rapidly on the same problem and for many problems this is the way to solve them. DS4 and GLM 5.3 shine when Qwen would take too many turns to solve something, they simply have more world knowledge and seem to home in on possible solutions faster in #tg but slower in wall clock time. If a particular problem is within striking range of all three then Qwen will usually be the fastest solver by far.
It also depends on how good your prompt is, whether you are setting up points for guidance or whether you expect everything to be solved in a single shot.
For me it has been a very mixed experience. I've had LLMs spit out 2K loc of relatively tricky code that worked the first time out of the gate. It compiled and ran and passed every test I threw at it. I've also seen them deadlocked on trivialities over 100's of toolcalls until I couldn't look at it anymore, stopped the session fixed it myself and continued from there. It's like having an idiot/savant with a weird mix of cleverness and sycophancy thrown in.
Sure. But the results are definitely better with better models.
GLM 5.3, Minimax, Sonnet, etc. I find are definitely worth using. They need handholding and prodding and you have to carefully review the output, but I almost always end up faster than doing it manually.
The GLM Quant is ~4 bpw, the DS4 is a Q8 and ninfer one is an NVFP4/BF16 I think. I run the ninfer one with ninfer-4090 on a hacked card with 48GB of VRAM. Max context for that one is 256K, DS4 has a 1M context and for GLM 5.3 it is 300K. The big models use all 8 GPUs.
Qwen benefits in two ways: all of the VRAM is on a single GPU so there is no communications other than generated tokens and new activations, that gives pretty good utilization for the card, the second is of course simply the fact that it is a much smaller model so there is less to compute. The 4090 is still power limited so it could do more but it's loud (and hot) enough as it is.
It strikes me that if it were up to Sam Altman, Eliezer Yudkowsky, Bill Gates and Dario Amodei, this crisis would persist indefinitely. If it were up to them I wouldn't be surprised to see them destroy the GPUs just them to keep them from the plebs.
I picked those names not to be hyperbolic, but because of very specific things I've seen them say.
Let us not forget, once they have all the chips they want, they can just sponsor a politician to start a war with China. Destroy semiconductor manufacturing capacity and they will have monopoly on compute for decades.
In fact I fail to see any other end game where they make profit on the huge liabilities that now have.
> ...the problem is the consensus for what comes after.
When one can't decide between two roughly equivalent options, there's a really cool way to choose between them! Assign each option to a side of a fair coin, flip the coin, and select the one whose side is face up.
It's really neat how many thorny problems can be either eliminated or made substantially less thorny with the careful addition of some randomness.
The sun rises when it rises, it has nothing to do with the clock. How does what the clock say take an hour of sunlight away from you? It is surely your employer doing that not the clock.
They're not equal. Summer time is awful in summer and it would be even worse in winter. We as a species need daylight in the morning and darkness in the evening. Not the other way around.
A new 3 series is wider, has a longer wheelbase, and approximately equal length of a 5 series from 2001. It's also ~500 lbs heavier, depending on options.
I drive a 1998 Lexus SC, a car that was built to compete with the contemporary 8-series and Mercedes SL. My car is comparable in size to a new Corolla. I'm longer, but I'm 4 inches shorter. Every other dimension is basically the same.
>The main reason kids get run over is unsafe crossing behaviors, such as running into the road after a ball. Older gets get run over because they're distracted by a smartphone.
(from the link)
Insane victim-blaming here. What is wrong with car people?
Agree, but in this case the new 2CV is shorter than the original and I imagine if the original was built now with today's safety and EPA requirements there wouldn't be that much difference
>It's ridiculous: Put a modern SUV next to a WW2 M4 Sherman and you will be surprised, the SUV is the same size. Some a bigger.
That would be what normal people would call a "bold faced lie" in normal conversation and is only above average levels of misleading compared to what's typical around here (which reflects upon the character of the people on HN, IMO).
A sherman is approx same length as the biggest modern SUVs but wider, taller and 10x the weight.
You may as well compare rowing shells to fishing vessels.
An M4 Sherman is about 6m long which production SUVs don't reach. That's about the length of an L3 delivery van, which is usually the longest version they make.
The M4 is 2.6m wide. No car is 2.6m wide. That's wider than a coach bus.
The M4 is 3m tall. Even the fattest SUVs do not come close.
My parents ID4 makes my old e39 touring look small by comparison. Funny thing is, it does not feel bigger on the inside, suppose modern safety standard have demanded a sizeable increase in vehicle size over time, and the battery bank also makes it grow on the, i suppose, x-axis
> But because the person next to you drives a big car and you don't want to get crushed, you also need a big car.
My position on this changed to "ah, now I see" after seeing a video where a Miata got run over from a standing start by a lifted truck. They were both stopped at the lights, but because the truck driver lacked both forward vision and object permanence, when the lights went green he set off .. directly over the car in front.
US size escalation will continue until legislated against. It's creeping into the UK, but fortunately you can't go too large until you start colliding with solid objects. And the max weight is limited to 3500kg unless you're on a Boomer license.
I think the trend towards SUVs was mainly because some car tax code incentivized it. If cars were taxed by weight (as they should - heavy vehicles cause so much more wear for public roads), the trend could reverse. (Plus you get better mileage with a lighter vehicle)
> If cars were taxed by weight (as they should - heavy vehicles cause so much more wear for public roads)
The problem with this argument is that because of the 4th power law, essentially all road wear due to vehicles is caused by 20+ ton commercial trucks and the difference between the different sizes of personal vehicle is immaterial.
If you want a tax that accurately reflects road wear it ends up being around $0.10/year for a 2500 pound subcompact, $1/year for a 4500 pound SUV and >$6500/year for a 20 ton commercial truck, assuming they all travel the same number of miles. And in practice the commercial truck will typically see ~10x more miles too.
A difference of $0.90/year in tax is probably not going to provide much incentive to anyone.
It is tax code and efficiency standards; SUVs are legally speaking light trucks and have lower fuel economy standards, so the Big 3 in the US just made their cars bigger rather than try to compete with the likes of Toyota and Honda on fuel efficiency.
The correlation between larger vehicles and profit margins is mostly because larger vehicles also cost more to operate, and then people with less money tend to avoid vehicles they can't afford to put gas in, so the people who buy larger vehicles usually have more money. In other words, nearly all large vehicles are premium vehicles whereas some small vehicles are the economy segment.
But that wouldn't stop them from making high margin vehicles that are low to the ground. The automaker with the highest margins is Ferrari.
The real problem is the nonsensical CAFE rules. If you want to make a high production volume small car with premium performance, you're going to blow your fuel economy target for that vehicle class, so they're de facto banned from making those. But >300 horsepower mid-sized SUVs? No problem. So premium performance cars are banned but premium performance "trucks" are fine, what are the premium buyers going to buy?
The companies are just making the thing people are buying. Pass stupid laws, get stupid cars.
A classic case of badly thought out rules/incentives getting perverse outcomes? But CAFE rules seem to be a US thing and we are seeing car size inflation in the UK, even if not to the same extent.
>The companies are just making the thing people are buying.
It is not clear to me to what extent the car companies are catering to the demand for big SUVs or creating the demand for big SUVs.
UK roads and parking spaces aren't getting any bigger. It is all going to end very badly.
If you see an old car next to a modern car, the size difference is extreme!
My daily driver is 220 inches long and 80 inches wide, weighs around 4000 lbs and over 50 years old. Most cars of that era were roughly the same size. It is definitely larger than most other cars I see today, although it also has much better visibility.
Most likely European vs American perspective? E.g. put an old and new Mini next to each other and it's almost unbelievable how tiny the original Mini actually is. While when you see a vintage US 'street cruiser' in a European city it's longer and wider than most modern cars.
Have a look at the original Fiat 500, it's comically small. You need to remember that the average adult Italian man in 1955 was about 8 cm shorter than today's average, and it's probably similar in other european countries.
Indeed - their original design language doesn't really scale. A bit like Ferrari and Porsche - their 4-door monstrosities scream mid-life crisis louder than a red convertible.
When interviewed about the new (BMW) Mini circa 2001, the designer of the original BMC Mini's hydrolastic suspension, John Moulton, said that calling it 'Mini' was ridiculous: "For heaven's sake, it's the size of an Austin Maxi!"
It would also likely result in your death if you hit anything solid. Cars are bulkier and visibility has been compromised by better crash protection. I'm reminded of that test they did years ago between a 1959 Impala and a 2009 Malibu. It did not go well at all for the Impala.
That was also a "specially engineered to show how far we've come" comparison. That generation of the impala was unconventional (in a bad way) for the time and considered particularly unsafe in a crash even by 1950s standards and they did a partial overlap with something much newer. They always do partial overlaps in these marketing videos (this is not a "test", they knew exactly what would happen when they spec'd it out) because that's where the biggest disparity is.
GM's X-frame cars were definitely weaker than average for the time, but going back to the original topic, I don't think a 2CV would stand a chance against one.
By the 70s, which is what I have, they were already much better, largely due to big shock-mounted bumpers and very long front ends that make for a gradual crumple zone:
That's an even more compelling video. Very small newer car vs a Volvo from the 90s, I would not have predicted that the Volvo would have done as poorly as it did. I totally expect cars from the 50s to basically disintegrate, but (showing my age) I had better opinions of 90s cars.
Yeah, that's why I said bulkier, I was trying to address the observation that visibility has gotten worse. I agree that it has, and a big reason is that A-pillars today are way wider than they used to be, because they are a critical part of the crash structure. They also typically have airbags, though in my experience those are compact and much narrower than the pillar itself.
I wonder if that difference is actually the safety standards rather than the efficiency ones though.
If you look at a modern vehicle, it's wide at the base and narrow at the top. The older ones were basically square, which is aerodynamically rubbish. But since the glass is at your eye level rather than below your waist, it meant you had a lot more glass.
I like to compare the 1960's Plymouth Suburban to the current Chevy Suburban both because of the name coincidence and because they cover roughly the same market segment (family hauler).
The Plymouth suburban is 9 inches shorter in length and well over 1000lbs lighter. I couldn't find specifications on height for the Plymouth, but the Chevy is obviously much taller.
They probably will be the one going through their parents stuff and hopefully by the time its an issue can move it to their kids' place. Options for trusted free hosting are limited sometimes. ;)
Sorry if it was not clear: The NAS is at my house but houses may burn down. THE discs are at my parents house for off-site backup. Like the sibling poster says, I would be the one going through the things anyway. But if unsure you can also ask a friend to keep them, one blu-ray can hold tons of photos and doesnt take much space.
Fair. Nothing is risk free. Move overseas, be more than 10 days away, other sibs have the burdens "we threw out those old unreadable backups Dad had stashed from his time at the computer center".. But your strategy isn't the worst by far. My 3-2-1 has 2 of the 3 in the same physical location. Not good.
Yes! For me, I like that foobar2000 (1) plays Ogg Vorbis files, and (2) will spin itself up as an FTP server on my LAN so I can easily upload music to it.
One difference between games and apps is the way of development and release. A game is build and then released to the community in one go, so everyone is assumed to be a learner.
Games are explicitly tested by new testers all the time to identify learning bottlenecks. App features are build for existing customers where the assumption is they already know how to use it, testing with new users seems to be rare.
Double down on this. I had a way worse back and bad pain in my early 20s than now later in life. I started with just walking an hour a day, later added gym. You don't need to become the next Arnold, just show up consistently.
One side note: There are core strength/posture issues but also issues that need medical attention. The tricky part is figuring out which.
Not about range, but speed. Wifi is surprisingly slow. I noticed downloading a model from the internet (Huggingface) to my Desktop (connected by LAN cable) is faster than copying it from said Desktop to the laptop in the same room via Wifi.
Yeah. I ran a hand crimped cat-5 Ethernet run to my office and speed tested it, and it turns out my wifi is faster, according to speed.cloudflare.com. Haven't had a chance to debug that yet, but there's a switch and some other wiring in the way.
Japan here and there is a word for it "nominication" for nomu (to drink) and communication. Instead of Christmas party we have a year end party (bonenkai). Otherwise reads the same. It's common and many people actually like it. There can be negatives, like the pressure to drink too much. But I never heard of anyone getting fired in all my social circle.
Of course the proprietary models will continue to improve and we have no guarantee that there will be new open models. But for coding, the local models are already good enough for a vast amount of use cases, if you are an experienced developer yourself and can guide it.
reply