npn

sorry LLM output is considered public domain in my country.

nah, both point to the same model for deepseek. it's just openrouter weirdness.

I don't think so. there is a lot of tools with similar usage, some harness even bring their own internal tools for accurately manipulation.

also, even if some models claim that they have full 1M context window, some only work effective with the head or tail of the window, a proper harness tool will know about the limitation of the model and act accordingly.

then also the output format, the tool calling syntax, the quirks and gotchas of each model.

it is not simple as just throwing everything at the model, especially when your project has hundred of files or so.

wait for Deepseek Harness (yes it is the official name) release then try again.

for your kind of task, harness tools matter.

I still believe this is not the full potential of pro models. I expect they will release another checkpoint later this year.

well, I think it is easier for the developers, we don't have to change the model name everytime. deepseek also does not have resource to serve 2 different v4 pro versions.

I think both are after qwen 3.8 max release.

thank you very much for the input. I once said to my friend that "no actually you do not get $20k worth of API usage for $200 subscription, it is mostly $600". turn out it was a pretty close guess.

even if anthropic does not apply all the fancy techniques that deepseek is using for their cache system, tiered kv cache system is something that anyone implements. the US labs even have huge CXL memory pool system. basically they charge us premium for what is almost free. no wonder they are pretty generous with the quota.

inb4 "but you will have to use api if not for the sub", no it is the other way around: because you have the subscriptions so people rarely have to pay for that input cache markup. if they have to pay that much money they would just simply switch models.

I suggest you make two other calculations:

1. Only calculate the output cost, skip input cost 2. Only calculate the cached input cost

You will be surprised

Agreed. I think people tend to forget that:

you do not have to pick a single language. it is trivial to use another language just for that single library. there are many way to communicate, it is not like method calling is always the hard requirement

since you do not have to choose a single language, the killer feature of that language is mostly useless, even if there is no other language that have the similar feature, people would just write a some glue code to bridge it with their main codebase. and usually, there is no such thing at unique feature that only that a single language can have. sure, it can be achieved easier in that language for some use case, but even so, a single bridge file is probably enough. aka nobody would invest in your language.

writing software is mostly using other people libraries anyway. the thing we write the most is some common business logic, thus the syntax is extra important, or rather, the DX is the most important thing. heck if the experience is great enough I can throw thousands hours to write the libraries/frameworks myself, it is (usually) not that hard.

it always turns me off seeing a new introduced language that syntax wise is a carbon copy of rust or C. no thank you, if you not even want to think about the syntax, why should we believe that you can deliver something truly innovative?

this is just survivorship bias. I bet you 1000 dollars that there exists hundred languages that are more useful and well thought than those languages you listed, but still nobody cared and it died.

I believe the reason lies entirely outside of the language itself. let just say, the is a circle, a society (of programmer and related peopel) that decide that a language can be popular or not. if your language get picked, then it will be popular, if not, then good luck.

you can deduce easily what the criteria (to be picked) from the list of languages you mentioned: 1. if it does not come from US, good luck (ruby is an exception but it didn't sail until some americans decide to boost it) 2. if the language authors do not speech english well, good luck (I mean, sure if you advertise your language more more people will pick it) 3. if you satisfy 2 previous conditions, if you do not play along well with those people in the circle, then good luck (D is a good example of this) 4. even your programming language is dog shit, as long as you know the people in the circle, they will promote it for you. bonus point: if you allow the language to be directed as they want, then it will be extra smooth.

tl;dr: it is mostly politic how a programming language gets popular.

weak argument. deepseek v4 flash is open weight, you can easily find other providers with competitive price with Deepseek (except for input caching), some even half as cheap.

From the private conversation of deepseek CEO and the investors, I'm under impression that they take pride for the low API price though. So it is not only training data. If they really want the data there are many other ways to do so.

btw, GA usually means it is suitable for production, and available globally.

the last time a gemini model was available globally in multiple datacenters was gemini 2.0 era.

Similar to aistudio vs vertex.

Or rather, Alibaba cloud is equivalent to google cloud or aws.

More like it is a generic idea that has been implemented 1000 times before so LLM already have the perfect solution.

Well, not like I'm saying the product is doomed though, because like always implementation details matter.

weird, for so long I'm damn sure that researchers are obligatory to release research papers to fulfill their quotas. otherwise they will lost their titles.

so unless those companies (openai, anthropic) gave the researchers some fortune big enough to make them give up their research careers, they will keep publish research papers. aka it is not the good will of those companies.

well usually you can just generate the data using LLM. use 2 or 3 different frontier models from different providers, then compare the results and pick the consensus. yes it is not 100% correct like when you make it manually, but then again even human makes mistakes.

There is a list of a few websites that I subconsciously think only have garbage articles and never really pay serious attention to them: quora, medium and substack.

Surely dev.to also filled with low level articles, but at least people there spent effort to write articles then gave it for free. But those users of the previous websites mostly write it for money.

Don't get me wrong, quality articles deserve to have money in return, you can't get far without sustainable business. But writing for the sake of writing (and earning money for the quantity not quality) is what is causing those websites to be filled with shallow contents.

fine tuning a small LLMs or even real small language models (like bert) is what the recommended way since the introduction of LLM. the benefit is pretty much obvious: faster to run, fully controlling the stack, better fit for the custom domain...

but in practice not many people do the fine tuning, for pretty much a single reason: large language models API cost are still very cheap, fast enough, and get improvement all the times. what the point of spend time (and money) to fine tune a specified model, then just when you release it a newer gen generic model is released and beat it?

but if someday the progress for LLM is slowed, or the price increased to the point calling api is not a viable approach any more, then surely the day of fine tuning and small models will come again.

even deepseek, with their current (dirt cheap) price, can earn enough profit to cover the cost (hardware investment?) in 10 months.

Huawei chips need advanced 3d packaging in order to keep up. Since the process is too complex the yield is still bad.

Not to mention there is a lot of demand from various factors, not deepseek only. Huawei itself is a major consumer.

They spend $40mil for lobbying, I'm sure they can also spare some millions to this place (and other places like reddit). They all do.

AI means it has npu, Max+ is marking the memory channel, PRO is a normal label for chips that have extra security baked in, it has been this way since forever.

I don't see anything wrong with the name.

update: the knowledge cut off date is "unknown" now.

funny because some people downvoted me believed that there is no relation between knowledge cut off date and real world events. that's not how it works!

tested the models on aistudio. despite that the knowledge cut off is march 2026 it still knows nothing about 2025!

you can check by asking "list notable world events in 2025, only list unplanned" on aistudio. or you can ask for Charlie Kirk, it also does not know. I tried it multiple time to ensure that I didn't not get routed to older models!

> but google has search

irrelevant, without deeper knowledge about cutting edge technologies or latest libraries, all of it suggestions are crap. even you ask it to search it will still use outdated keyword thus only getting outdated information.

in other word, what a disaster!

what a horrible article. full of misinformation and dishonesty.

1. training new base models are expensive for sure, but fine-tuning them are relatively inexpensive enough the labs can continue to do so forever. the main reason why frontier models are so good is because the massive input they generated from user usage. they are using that information to strategically build better training data. and this is why no other models can catch up, til now that is. but if chinese models are good enough, and free to host, and cheaper to use, then the consequence is the frontier labs will lost valuable user inputs and the chinese labs will gain more. as time goes by this will be a domino effect.

2. nvidia is not only the player in the hardware scene. amd mi350p is getting popular, and huawei is pumping SuperPoDs. what does this mean for us? chinese models will surely use chinese hardware, and optimize for them. the other people will pick amd because compare to nvidia they are cheaper. with open weight models and open source inference stacks, they are freely to experiment and improve the stack, thus further lower the inference cost and nvidia dependency. and they even plan to build their own inference hardware, too. and nvidia loses market share meaning all the fund it gives to openai or anthropic will be cut, too.

and you say there is nothing to afraid?

Did they cry about it? No right? Don't apply your own standard then judge them about it, petty people.

Stackoverflow aimed to be a knowledge base. And knowledge base has a ceiling limit. They simply reached the point that almost all questions (regarding the knowledge) were asked for them. You can argue that newer or niche libraries or languages knowledge is still lacking there, but I have never seen them getting closed, just not answered.

Not worth it. I have just tried a single prompt in the web interface and it is still not finish reasoning. It thinks too much and often repeats the same stuff over and over.

Combine with the price it will surely more costly than gpt 5.6.

it is funny because nobody ever bother points out that they overcharge you for text input token price.

sure it was pretty resource intensity a few years before, but with turbo quant, sparse attention and various techniques, plus the advancing of hardware (dedicated prefill machine, memory pool for kv caching) the cost should be drastically reduced, and yet they still keep the same cost formula.

I can't help but laugh whenever someone proudly share how many billion input tokens they spent in their code sections and how much they saved with the subscription, meanwhile it is pretty much just electricity cost for the providers.