Hacker News

Still not sure how I feel about China of all places to control the only alternative AI stack, but I guess it's better than leaving everything to the US alone. If China ever feels emboldened enough to go for Taiwan and the US descends into complete chaos, the rest of the world running on AI will be at the mercy of authoritarian regimes. At the very least you can be sure noone is in this for the good of the people anymore. This is about who will dominate the world of tomorrow. And China has officially thrown their hat in the ring.

As much I apprecite the sentiment, I think it is too early to declare that the well guareded monopoly is over. Yes, these models have answers, but don't expect all the large enterprises to switch to these models. The other aspect is scaling to serve these models will need a lot of time even if Huawei succeeds. Not all the Governments trust China and there will be a lot of resistance to work with these models eventually, even if cheaper.

I've been baffled watching America double down on the same strategy even when it failed to produce results

They sanctioned the hell out of Huawei and now Huawei is bigger than ever

America is just not able to digest the idea that another country can be as good, if not better, at innovation

not really, china has gone domestic for everything as soon as it could.

its naive to think they would have stayed on a 'western' stack.

Most of the time 'losing' isn't making a bad choice its being put in a situation where you have no good choices.

I don’t know if we’re ahead of the curve but that tired feeling has started turning into hate here in the EU. I guess being threatened with invasion does that to you.

The next decade is going to look very different with America Alone.

"not perfect" is a _very_ big simplification of what China is though

This is such a tired argument, and morally repugnant. Where is the UK in the race, where is the EU? Lets get of our asses and stop moralizing.

(China wiped out the entire EU industry through a "quiet" trade war since like the last 15 years, and we're not really talking about that aren't we...)

america is a continent. let’s take back our vocabulary (fellow european here). the little orange man shows very well what i mean when he started giving names to the gulf of mexico.

As someone that lived in Britain for 15 years until 2024, I'm not sure a nation with a GDP per capita lower than Poland, that is now poorer than every state in America, with a gang rape epidemic the government tried to suppress investigating should really concern itself with how other countries are ran.

As a different Brit I do not accept such moral relativism.

China’s governments actions are on a completely different level - for example:

“””

Since 2014, the government of the People's Republic of China has committed a series of ongoing human rights abuses against Uyghurs and other Turkic Muslim minorities in Xinjiang which has often been characterized as persecution or as genocide.

“”” https://en.wikipedia.org/wiki/Persecution_of_Uyghurs_in_Chin...

https://www.amnesty.org/en/location/asia-and-the-pacific/eas...

Yes Trump is clearly trying Totalitarianism in America, but it is orders of magnitude different from what is happening in China.

Fellow countryman here. I came here to say the same thing

If you don't have the source code then it makes no difference. If you have the weights and are running some model via llama.cpp, then you are using whatever API llama.pp is using, not the API that was used to train the model or that anyone else may be using to serve it.

If you found a rare 9000 card with 200+ GB of VRAM, sure

> Let's see how long it takes before the big US AI companies start lobbying to outright ban use of Chinese AI, even the open source / local models. For "national security" reasons, of course.

Already do on EVs.

Hopefully the US’ self imposed isolation will mean that when they do, they aren’t able to force the rest of the world to follow suit.

>I asked DS itself and it denied this

Bro, seriously?

This is a pretty banal comment at this point. Open source is the term used in the LLM community. It's common and understood. Nobody is going to release petabytes of copyrighted training data, so the distinction between open source vs weights is a rather pointless one.

For me open source means that the entire training data is open sourced as well as the code used for training it otherwise it's open weight. You can run it where you like but it's a black box. Nomic's models are good example of opensource.

you talk like there isn't censorship in american AIs, like Israel topics.

I can't wait for Taiwan to peacefully reunify with the mainland so the west with its constant war waging won't even have this talking point

> China asks other country not to meddle with internal separatism > They also dont support separatism in my country

Understandable.

Just ask it for a summary of the USA’s role in Iran, Gaza, Lebanon and its recent threats against Panama, Cuba and Greenland! It might be able to keep track.

Quit a bit better then made to bomb little girl schools in Iran.

pretty sure you can ask whatever you want and it will tell you official stance agreed by almost all countries in the world that Taiwan is part of China as it's recognized by your own country (I don't even know where are you from, but there is like 98% chance I'm right)

Theoretically yes. It is entirely possible to poison the training data for a supply chain attack against vibe coders. The trick would be to make it extremely specific for a high value target so it is not picked up by a wide range of people. You could also target a specific open source project that is used by another widely used product.

However there is so many factors involved beyond your control that it would not be a viable option compared to other possible security attacks.

If there is, couldn't they exist in any model?

I don't mean that flippantly. These things are dumped in the wild, used on common (largely) open source execution chains. If you find a software exploit, it's going to affect your population too.

Wet exploits are a bit harder to track. I'd assume there are plenty of biases based on training material but who knows if these models have a MKUltra training programme integrated into them?

Backdooring software at scale.

Spearphishing.

Building reliance and exploiting it, through state subsidies, dumping, and market manipulation.

Handicapping provision to the west for competitive advantage.

What about LLMs from other origins? What makes them less risky?

From my experience, kinda the opposite? It's like Chinese software is... Harder to weaponize or hurt yourself on. Deepseek is definitely censored, but I've never caught it being dishonest in a sneaky way.

There must be. The executives at my company wouldn't have banned them all for no reason after all.

Is this a serious comment? It honestly reads like the last famous words.

Of course there are risks.

Very cool to hear there is agreement with (probably quite challenging?) coding problems as well.

Just ran a couple of them through GPT 5.5, but this is a single attempt, so take any of this with a grain of salt. I'm on the Plus tier with memory off so each chat should have no memory of any other attempt (same goes for other models too).

It seems to be getting more of the impressive insights that Gemini got and doing so much faster, but I'm having a really hard time getting it to spit out a proper lengthy proof in a single prompt, as it loves its "summaries". For the random matrix theory problems, it also doesn't seem to adhere to the notation used in the documents I give it, which is a bit weird. My general impression at the moment is that it is probably on par with Gemini for the important stuff, and both are a bit better than DeepSeek.

I can't stress how much better these three models are than everything else though (at least in my type of math problems). Claude can't get anything nontrivial on any of the problems within ten (!!) minutes of thinking, so I have to shut it off before I run into usage limits. I have colleagues who love using Claude for tiny lemmas and things, so your mileage may vary, but it seems pretty bad at the hard stuff. Kimi and GLM are so vague as to be useless.

I don't want to give away too much due to anonymity reasons, but the problems are generally in the following areas (in order from hardest to easiest):

- One problem on using quantum mechanics and C*-algebra techniques for non-Markovian stochastic processes. The interchange between the physics and probability languages often trips the models up, so pretty much everything tends to fail here.

- Three problems in random matrix theory and free probability; these require strong combinatorial skills and a good understanding of novel definitions, requiring multiple papers for context.

- One problem in saddle-point approximation; I've just recently put together a manuscript for this one with a masters student, so it isn't trivial either, but does not require as much insight.

- One problem pertaining to bounds on integral probability metrics for time-series modelling.

>Western Models are optimizing to be used as an interchangeable product.

But so much investment in their platforms, not just their APIs?

> Western Models are optimizing to be used as an interchangeable product

Why? It sounds like the stupidest idea ever. Interchangeability = no lock-in = no moot.

it's in the footnote text of the first figure of the section the link points to, where "昇腾950" means "Ascend 950"

Research and training costs have to be amortized from somewhere; and labs are always training. I'm definitely keen for the financials when the two files for IPO though, it would be interesting to see; although I'm sure it won't be broken down much.

In six month deepseek won't be sota anymore und usage will be wayyyy down.

Amazingly, the current depreciation overestimates the retained value of GPUs.

In 2023, the depreciation schedule for H100s was 2 years, but they are still oversubscribed and generating signficant income.

Coreweve has upped their depreciation for GPUs to 6 years(!) now, which seems more realistic.

https://www.silicondata.com/blog/h100-rental-price-over-time

I'm pretty sure OpenAI and Anthropic are overpricing their token billed API usage mainly as an incentive to commit to get their subscriptions instead.

Before the AI bubble that will burst any time now, there was the AI winter that would magically arrive before the models got good enough to rival humans.

The claim that APIs are subsidized is very common.

eg:

Token prices are significantly subsidized and anyone that does any serious work with AI can tell you this.

https://news.ycombinator.com/item?id=47684887

(the claims don't make any sense, but they are widely held)

Yeah, subscriptions used to be extraordinarily generous. I miss those days, but the reinvigoration of open weight models is super exciting.

I'm still playing with the new Qwen3.6 35B and impressed, now DeepSeek v4 drops; with both base and instruction-tuned weights? There goes my weekend :P

Google stated a while back, that with tpus they are able to sell at cost / with profit.

Aka: everyone who uses Nvidia isn't selling at cost, because Nvidia is so expensive.

These models are open and there are tons of western providers offering it at comparable rates.

As this is a new arch with tons of optimisations, it'll take some time for inference engines to support it properly, and we'll see more 3rd party providers offer it. Once that settles we'll have a median price for an optimised 1.6T model, and can "guesstimate" from there what the big labs can reasonably serve for the same price. But yeah, it's been said for a while that big labs are ok on API costs. The only unknown is if subscriptions were profitable or not. They've all been reducing the limits lately it seems.

you know, if you don't have to pay insane salary for your top engineers, and don't have to pay billions for internet shills to control the narrative, then all of the labs will be insane profitable.

When China does it it's communism. When companies in the west get massive tax cuts, rebates, incentives and subsidies, that's just supporting the captains of industry.

doesn't it get tiring after a while? using the same (perceived) gotcha, over and over again, for three years now?

no one is ever going to release their training data because it contains every copyrighted work in existence. everyone, even the hecking-wholesome safety-first Anthropic, is using copyrighted data without permission to train their models. there you go.

They are exactly open source. The training data is the internet. Don't say it's on the internet. It IS the internet.

The training scripts are in Megatron and vLLM.

Aww yes, let me push a couple petabytes to my git repo for everyone to download...

Weights are the source, training data is the compiler.

This is only good advice if you don’t have the need to understand what’s happening on the edge of the frontier. If you do, then you’ll lose on compounding the knowledge from staying engaged with the major developments.

It is funny seeing people ping pong between Anthropic and ChatGPT, with similar rhetoric in both directions.

At this point I would just pick the one who's "ethics" and user experience you prefer. The difference in performance between these releases has had no impact on the meaningful work one can do with them, unless perhaps they are on the fringes in some domain.

Personally I am trying out the open models cloud hosted, since I am not interested in being rug pulled by the big two providers. They have come a long way, and for all the work I actually trust to an LLM they seem to be sufficient.

I didn't express this well but my interest isn't "who is in the top spot", and is more _why and _how various labs get the results they do. This is also magnified by the fact that I'm not only interested in hosted providers of inference but local models as well. What's your take on the best model to run for coding on 24GB of VRAM locally after the last few weeks of releases? Which harness do you prefer? What quants do you think are best? To use your sports metaphor it's more than following the national leagues but also following college and even high school leagues as well. And the real interest isn't even who's doing well but WHY, at each level.

For coding Opus 4.5 in q3 2025 was still the best model I've used.

Since then it's just been a cycle of the old model being progressively lobotomised and a "new" one coming out that if you're lucky might be as good as the OG Opus 4.5 for a couple of weeks.

Subjective but as far as I can tell no progress in almost a year, which is a lifetime in 2022-25 LLM timelines

On-device is incredibly far away from being viable. A $20 ChatGPT subscription beats the hell out of the 8B model that a $1,000 computer can run.

Nvidia's forward PE ratio is only 20 for 2026. That's much lower than companies like Walmart and Costco. It's also growing nearly 100% YoY and has a $1 trillion backlog.

I think Nvidia is cheap.

Try interacting with it through the website, it will give an error and some explanation on the issue. I had to relax my guardrail settings.

Its on OR - but currently not available on their anthropic endpoint. OR if you read this, pls enable it there! I am using kimi-2.6 with Claude Code, works well, but Deepseek V4 gives an error:

`https://openrouter.ai/api/messages with model=deepseek/deepseek-v4-pro, OR returns an error because their Anthropic-compat translator doesn't cover V4 yet. The Claude CLI dutifully surfaces that error as "model...does not exist"

I tried OpenRouter but I feel the money flies even with these models, it is not comparable to a subscription but yes, it's very good for trying. Maybe I should test other models alongside GPT 5.5 to see which one fits me.

I am thinking about getting some 1 year promotion as a student before defending my PhD.

I just want to try to make something useful out of my time, that's why I'm subscribed to Codex at the moment. 20€ is affordable, not really a problem. But yes, maybe I would do me a favor unsubscribing and going back to the old ways to learn properly.

I had Claude make me a quick tool to combine my Claude Code token usage (via ccusage util) with OpenRouter pricing from the models API

I'm on Max x5 plan and any of the 'good' models like Kimi 2.6, GLM, DeepSeek would have cost 3-5x in per-token billing for what I used on my Claude plan the last three months

So unless my Claude fudged the maths to make itself look better, seems like I'm getting a good deal

I am not so sure, credits fly when using any model trough API if I use it as much as I use Codex.

It is tricky to build good infrastructure for prompt caching.

It’s not remotely hypothetical you’d have to be living under a rock to believe that. And the fusion with a one-party state government that doesn’t tolerate huge swathes of thoughtspace being freely discussed is completely streamlined, not mediated by any guardrails or accountability.

This “no harm to me” meme about a foreign totalitarian government (with plenty of incentive to run influence ops on foreigners) hoovering your data is just so mind-bogglingly naive.

As someone with Tibetan friends and as someone from India, Chinese ulterior motives are way more clear.

Same

> Internet comments say that open sourcing is a national strategy, a loss maker subsidized by the government. On the contrary, it is a commercial strategy and the best strategy available in this industry.

This sounds whole lot like potatoh potahto. I think the former argument is very much the correct one: China can undercut everyone and win, even at a loss. Happened with solar panels, steel, evs, sea food - it's a well tested strategy and it works really well despite the many flavors it comes in.

That being said a job well done for the wrong reasons is still a job well done so we should very much welcome these contributions, and maybe it's good to upset western big tech a bit so it's remains competitive.

Please don't slander the most open AI company in the world. Even more open than some non-profit labs from universities. DeepSeek is famous for publishing everything. They might take a bit to publish source code but it's almost always there. And their papers are extremely pro-social to help the broader open AI community. This is why they struggle getting funded because investors hate openness. And in China they struggle against the political and hiring power of the big tech companies.

Just this week they published a serious foundational library for LLMs https://github.com/deepseek-ai/TileKernels

Others worth mentioning:

https://github.com/deepseek-ai/DeepGEMM a competitive foundational library

https://github.com/deepseek-ai/Engram

https://github.com/deepseek-ai/DeepSeek-V3

https://github.com/deepseek-ai/DeepSeek-R1

https://github.com/deepseek-ai/DeepSeek-OCR-2

They have 33 repos and counting: https://github.com/orgs/deepseek-ai/repositories?type=all

And DeepSeek often has very cool new approaches to AI copied by the rest. Many others copied their tech. And some of those have 10x or 100x the GPU training budget and that's their moat to stay competitive.

The models from Chinese Big Tech and some of the small ones are open weights only. (and allegedly benchmaxxed) (see https://xcancel.com/N8Programs/status/2044408755790508113). Not the same.

Weights are the source, training data is the compiler

I think their motives are pretty transparent, as are china’s, as ever, you have to pick the lesser of two evils.

Which model's best depends on how you use it. There's a huge difference in behaviour between Claude and GPT and other models which makes some poor substitutes for others in certain use cases. I think the GPT models are a bad substitute for Claude ones for tasks such as pair-programming (where you want to see the CoT and have immediate responses) and writing code that you actually want to read and edit yourself, as opposed to just letting GPT run in the background to produce working code that you won't inspect. Yes, GPT 5.4 is cheap and brilliant but very black-box and often very slow IME. GPT-5.4 still seems to behave the same as 5.1, which includes problems like: doesn't show useful thoughts, can think for half an hour, says "Preparing the patch now" then thinks for another 20 min, gives no impression of what it's doing, reads microscopic parts of source files and misses context, will do anything to pass the tests including patching libraries...

Agree with your assessment, I think after models reached around Opus 4.5 level, its been almost indistinguishable for most tasks. Intelligence has been commoditized, what's important now is the workflows, prompting, and context management. And that is unique to each model.

Is Opus nerfed somehow in Copilot? Ive tried it numerous times, it has never reallt woved me. They seem to have awfully small context windows, but still. Its mostly their reasoning which has been off

Codex is just so much better, or the genera GPT models.

I found Opus 4.7 to be actually worse than Opus 4.6 for my use case

Substantially worse at following instructions and overoptimized for maximizing token usage

This resonates with me a lot.

I do some stuff with gemini flash and Aider, but mostly because I want to avoid locking myself into a walled garden of models, UIs and company

What do you run these on? I've gotten comfortable with Claude but if folks are getting Opus performance for cheaper I'll switch.

actually this is not the reason - the harness is significantly better. There is no comparable harness to Claude Code with skills, etc.

Opencode was getting there, but it seems the founders lost interest. Pi could be it, but its very focused on OpenClaw. Even Codex cli doesnt have all of it.

which harness works well with Deepseek v4 ?

eh idk. until yesterday opus was the one that got spatial reasoning right (had to do some head pose stuff, neither glm 5.1 nor codex 5.3 could "get" it) and codex 5.3 was my champion at making UX work.

So while I agree mixed model is the way to go, opus is still my workhorse.

I don’t know what people are doing but Minimax produced 16 bugreports which of 15 was false positives (literally a mistake).

In contrast ChatGPT 5.3 and also Opus has a 90% rate at least on this same project. (Embedded)

All other tests were the same. What are you doing with these models?

The thing is, it doesnt need to beat 4.7. it just needs to do somewhat well against it.

This is free... as in you can download it, run it on your systems and finetune it to be the way you want it to be.

Tbh I was more productive with 4.6 than ever before and if AI progress locks in permanently at 4.6 tier, I’d be pretty happy

It is more than good enough and has effectively caught up with Opus 4.6 and GPT 5.4 according to the benchmarks.

It's about 2 months behind GPT 5.5 and Opus 4.7.

As long as it is cheap to run for the hosting providers and it is frontier level, it is a very competitive model and impressive against the others. I give it 2 years maximum for consumer hardware to run models that are 500B - 800B quantized on their machines.

It should be obvious now why Anthropic really doesn't want you to run local models on your machine.

Apparently glm5.1 and qwen coder latest is as good as opus 4.6 on benchmarks. So I tried both seriously for a week (glm Pro using CC) and qwen using qwen companion. Thought I could save $80 a month. Unfortunately after 2 days I had switched back to Max. The speed (slower on both although qwen is much faster) and errors (stupid layout mistakes, inserting 2 footers then refusing to remove one, not seeing obvious problems in screenshots & major f-ups of functionality), not being able to view URLs properly, etc. I'll give deepseek a go but I suspect it will be similar. The model is only half the story. Also been testing gpt5.4 with codex and it is very almost as good as CC... better on long running tasks running in background. Not keen on ChatGPT codex 'personality' so will stick to CC for the most part.

Their Chinese announcement says that, based on internal employee testing, it is not as good as Opus 4.6 Thinking, but is slightly better than Opus 4.6 without Thinking enabled.

Anthropic fans would claim God itself is behind Opus by 3-6 months and then willingly be abused by Boris and one of his gaslighting tweets.

LMAO

FWIW it's also like 10x cheaper.

There appears a flight of dragons without heads. Good fortune.

That's literally what the I Ching calls "good fortune."

Competition, when no single dragon monopolizes the sky, brings fortune for all.

Output is what the compute is used for above all else; costs more hardware time basically than prompt processing (input) which is a lot faster

input tokens are processed at 10-50 times the speed of output tokens since you can process then in batches and not one at a time like output tokens

How can they fix it after the release? They would have to retrain/finetune it further, no?

The pro pelican is a work of art! It goes dimensions that no other LLM has gone before.

yeah. look at these 4 feathers (?) on his bum too.

a lot of dumplings

I tried that for the GPT-5 launch - a self-improving loop that renders the SVG, looks at it and tries again - and the results were surprisingly disappointing.

I should try it again with the more recent models.

they're just late for work.

996 Pelican, lol

This is an excellent comment. Thanks for this - I've only ever thought about whether the frame is the right shape, I never thought about how different illustrations might map to different bicycle categories.

The Pedersen looks like someone failed the "draw a bicycle" test and decided to adjust the universe.

It was a bigger deal in the Gemini 3.1 launch: https://x.com/JeffDean/status/2024525132266688757

>Generate an SVG of a pelican riding a bicycle

at the top of the linked pages.

https://news.ycombinator.com/item?id=47879092#47880421

In the 5.5 topic

This should be the bottom comment on the pelican comment on every model release post.

The sheer number of bugs and lack of meaningful improvements in Google products is a clear counterargument to the AI bull thesis

If AI was so good at coding, why can’t it actually make a usable Gemini/AI Studio app?

Most of these tests are one-prompt in nature. I've also noticed issues with the PDF reader in Gemini which was very frustrating, although it is significantly better now than it was even two weeks ago. On the contrary, now GPT-5 seems to be giving me issues.

In my experience, Gemini is the most insightful model for hard problems (particularly math problems that I work on).

I gave up on Gemini 3.1 Pro in VSCode after 2 hours. They fully refunded me.

Yeah if I could use Gemini with pi.dev that would be my choice. But Gemini CLI is just so, so bad.

Kimi is indeed somewhat cheap for frontier-level intelligence, but still is $4-5 per mm tokens. Deep Seek is at least an order of magnitude cheaper.

Because they all (if my memory serves) did this release at the same time thing last time. I have not looked into it but I am guessing that not letting one model pull ahead for a month means everyone keeps up - which implies the “stickiness” of any one model is a lot lower than we think

You can, but does it work well? I assume CC has all kinds of Claude specific prompts in it, wouldn't you be better with a harness designed to be model agnostic like pi.dev or OpenCode?

there's no pinging. Someone's gotta email dang

That's like saying Raytheon would outsource building drones from Saheed makers (don't know who exactly).

Not gonna happen

So, dual RTX PRO 6000

Jordan Keller worked at Moonshot? Or am I missing something? I thought he is the original author. https://x.com/kellerjordan0/status/1842300916864844014

Flash is less than 160 GB. No need to quantize to fit in 2x 96 GB. Not sure how much context fits in 30 GB, but it should be a good amount.

Very carefully

OpenCode? Pi?

If one finds it difficult to set up OpenCode to use whatever providers they want, I won't call them 'dev'.

The only real friction (if the model is actually as good as SOTA) is to convince your employer to pay for it. But again if it really provides the same value at a fraction of the cost, it'll eventually cease to be an issue.

You can literally run it from Claude code. Easily too

They have instructions right on their page on how to use claude code with it.

More like he wants to ban accelerator chip sales to China, which may be about “national security” or self preservation against a different model for AI development which also happens to be an existential threat to Anthropic. Maybe those alternatives are actually one and the same to him.

"2 minutes ago" https://huggingface.co/unsloth/DeepSeek-V4-Pro

Those are quants, not distills.

I've been calling that the "streaming experts" trick, the key idea is to take advantage of Mixture of Expert models where only a subset of the weights are used for each round of calculations, then load those weights from SSD into RAM for each round.

As I understand it if DeepSeek v4 Pro is a 1.6T, 49B active that means you'd need just 49B in memory, so ~100GB at 16 bit or ~50GB at 8bit quantized.

v4 Flash is 284B, 13B active so might even fit in <32GB.

These are more like experiments than a polished release as of yet. And the reduction in throughput is high compared to having the weights in RAM at all times, since you're bottlenecked by the SSD which even at its fastest is much slower than RAM.

Do you have the links for those? Very interested