Hacker News

I don't think you're taking the good-faith reading of the sentence's grammar here, which is:

"[When China is mentioned] the comments section stops discussing [technical aspects], and instead starts going on about [non-technical aspects], and all that [stuff which I do not feel is relevant]."

I'm struggling, though, to find a good-faith reading of what you meant by:

"Denying human rights. Classic. Sorry - you instantly lost any respect I could have maybe had for your opinion."

Specifically, "classic" - classic what? "Classic" meaning "a trait you personally believe is more prevalent in people who (might be) citizens of a specific country"?

Hopefully the good-faith interpretation of GP's comment, which you might have unintentionally missed, will help you engage with their opinion with more respect.

I think it's an easy opinion to empathise with personally. I would be enormously frustrated myself, on an individual level, if interesting technical achievements made by companies in my own country were overshadowed by political discussion about the actions of the country overall, when spoken about on technical forums.

This would be especially true for me if I - as an individual - disagreed with politically with my country's leadership. Frustration at a topic of discourse is not mutually exclusive with being in favour of human rights.

I don't think GP has done themselves any favours with the broad statements that close their comment, and is clearly frustrated themselves. That there's mutual frustration is probably a sign that empathy on both sides might lead to a really productive discussion - but focusing on the more substantive points they've made, rather than escalating by misinterpreting the less important aspects of what they commented - is what will lead to that.

You are not allowed to discuss human rights until your government stops its unprovoked war on Iran, issues a full repudiation of the policies that led to it, and makes full reparations to the entire rest of the world whose energy prices you have driven up for the benefit of a handful of wealthy inside traders.

Perhaps whenever a US firm, eg. Open AI, Anthropic, Gemini, Grok etc. release a model.

Top comments should be criticising US torture camps in Guantanamo bay or their recent bombing of a girls school in Iran.

> Denying human rights.

Taking words out of context. Classic. Sorry - you instantly lost any respect I could have maybe had for your opinion.

he never said anything about denying human rights

I don't see anyone talking about Trump, ICE executing people in the streets, the insane level of US politics, the slow build up of the surveillance state, etc when discussing claude or openai

I think he means rubbish as in completely off topic.

It's probably in Chinese, which would cut out most of us here.

Legally gray practices? Such as what?

Old politicians who barely understand how to use e-mail had a panic attack due to Anthropics marketing about Mythos and being told that Chinese models were "attacking" (distillation) US frontier labs to train theirs.

They pull the only lever they could in an attempt to stop it, but again, they don't understand how any of it works. The export ban lasted a few weeks yet everyone keeps crying about it like the US is shady.

In reality the entire world is 100% dependent on the output of US and China for frontier AI. Anyone outside of those two countries will be loud about either closing off access. I personally think it is inevitable that China and the US stops producing frontier level open models when they become too valuable.

Rankings have never been something I’ve cared about. Just as with the smear campaigns against China, there is far too much noise in the world. We simply hope that the geeks who are genuinely committed to making the world a better place can unite and focus on getting things done, rather than being brainwashed by ideology and social media.

> kinda wild that alibaba of all companies is making it hard to give them money tho

Okay so this is entirely off topic but, are there any options to subscribe to these companies without having a credit card? (I am Brazilian if this makes a difference)

It's kind of wild that a Chinese AI lab will require something like.. a credit card with visa or mastercard label, and very little else.

It's an interesting example of the metric becoming a target. Simon started drawing pelicans because it wasn't something that existed before.

In my opinion, simonw shouldn't have to play those games

> It's hard to say what their motivation is.

Feels pretty easy to me.

They want to turn LLMs into a commodity, and watch the US AI labs crash and burn.

There will still be plenty of customers who will pay them to host the models and run inference, even if the weights are open and others can offer competing products. (If necessary, the Chinese government can ban use of foreign inference services by Chinese citizens and businesses to give their own companies a domestic monopoly.)

When their models equal or surpass those from the Western AI labs, they can even stop releasing weights for new models, and keep all the inference revenue for themselves.

Meanwhile, they're still manufacturing much of the hardware that everyone in the world needs in order to run datacenters (see also: Spolsky's "commoditize your complement" essay).

Beyond that, it's a soft-power play. As the world keeps looking at the US more and more skeptically as an ally and superpower, Chinese companies releasing weights for competitive models is a way for China to look better and more world-minded.

In China, you can’t officially use US APIs. The world saw a taste of this with Fable, but in China, this has been the situation all along.

So it’s not a surprise why open weights are so cherished. As frontier models continue to block everyday individuals from securing their own codebase, I expect the adoption and usage of open weights to continue.

As an example, HuggingFace recently was investigating a security incident and got locked out of frontier closed APIs. Yes, HuggingFace.

https://huggingface.co/blog/security-incident-july-2026

> It's hard to say what their motivation is.

Not for anyone who reads history.

Back in the late 18th century, England was the world's top economy, in big part due to its textile industry. England had an export ban on the technology, but textile worker named Samuel Slater brought blueprints over (Supposedly in response to a bounty posted in a newspaper by the US government!). The technology diffused rapidly because the legal environment made competition easy, and ironically the US had better sources of energy (superior water-power sites).

Arguably, China is doing the same thing in the 21st century.

Fwiw, American industry has given away a lot for free - you could include large parts of the open source movement in that - and all the "free" VC backed services like facebook would be another prong of the same comparison. I would rather compare this way, that China is gaining soft power and goodwill, in the technology and innovation sense, in a way that's similar to how USA has done in the past.

There’s a Twitter thread making rounds by Dean Ball about deceleration in AI development caused by open models and I can’t understand how people don’t see that it’s true: open models dismantle the frontier lab capex spend potential by reducing the training budget to zero in the limit. Tokens from different providers are not fungible, but customers are nevertheless very price sensitive and close enough is good enough, eg. K3 being opus+ in capability and cheaper than opus per successful task in the long run is an obvious financial decision.

No training budget means deceleration, or at least slower acceleration, margin compression and a completely demolished IPO valuation; path to machine god requires dollars and capable open models externalize training costs to true frontier labs parasitically.

IMHO humanity has a better chance at not destroying itself due to less than breakneck pace - but there’s a chance frontier models get sponsored by the USG and are never released publicly so they can’t be distilled and then what?

> It's hard to say what their motivation is.

Not that hard to say IMO, they basically see models becoming a commodity and see value in the applications on top of them. So if Alibaba Cloud is the best place to build applications on top of Qwen, why not give the model itself away?

China is watching world sentiment shifting away from USA. Doing many small things that show both strength and openness is surely very intentional.

> It's hard to say what their motivation is

Why is it hard? Their government has been very clear that they plan to win on manufacturing: https://english.www.gov.cn/news/202601/08/content_WS695f1b55...

Technically they've been saying it for the last 40 years.

Popular open-source projects:

Google: Chromium, Kubernetes, Android, TensorFlow

Meta: React, PyTorch, Llama

Microsoft: VS Code, TypeScript, .NET Core

LinkedIn: Kafka

Slotted in along these, an analogous explanation is that Alibaba needs Qwen internally (vs depending on an American company), but licensing is not part of their revenue strategy. (As a cloud vendor, they can make money on inference. The strategy is very similar to the US hyperscalers ex-Google.)

Joel Spolsky wrote in depth about this notion of commoditizing one's complement in 2002[1] using tech examples stretching back into the '80s.

1 - https://www.joelonsoftware.com/2002/06/12/strategy-letter-v/

Humanity is a bit of a stretch, and to be seen over time, not that I'm saying it won't happen; let's get some hubris here.

In case of American labs, always follow the money. In case of Chinese labs, always follow the IP.

They are trying to make money. That's what firms in any capitalistic economy care about the most. Regardless of the government's presumed interference, the companies themselves are all trying to make money. All competing for subscriptions and API payments.

One aspect of this is making a name for yourself i.e. PR. Making a capable model open source helps a lot with that.

> It's hard to say what their motivation is.

Maybe because the industry isn't yet very sure as to what the use cases might be for these technologies they're hoping that by making it open source and accessible to everyone that someone could find interesting applications for it and even more so, perhaps, way to further the technologies themselves.

There are more Chinese than Americans, so statistically speaking, I'm guessing, there'd be a greater chance for one of Chinese engineers to make advancements than one of American. But that's pure speculation on my part, being neither, I'm just happy I can be a part of it and play with the tools as well~

Xi Pitches China as Leader of New Global AI Order, Challenging US Dominance:

https://www.reuters.com/world/asia-pacific/chinas-xi-promote...

> it also happens to be really good for humanity.

AI being good for humanity is still an open question, but for closed vs. open models/weights, yeah it is preferred. I foresee it won't be much longer before everyone will be slicing/distilling/tuning their models once the architecture improves.

The Chinese firms may just be making a bad business decision.

I'm paraphasing but the Chinese premier said recently AI should be seen as a common good that should benefit everyone.

Its about closing the gap. its the gap over everyone else that will give one country leverage over everyone else in the AI age. Makes me wonder what the world would look like if a country or group of countries did this during the industrial revolution.

This is their motivation https://try.works/why-chinese-ai-labs-went-open-and-will-rem...

It’s too soon to say if it’s good for humanity; that might be overly optimistic. Commodity markets aren’t always good (for example, arms or drug markets). Will LLM’s turn out like one of those? There are people I respect arguing in favor of more regulation.

What I understand is that by doing this it seems like profit will shift to chip makers,as we'll run more models locally, and currently American companies have the advantage here.

So what would the long game be for chinese companies?

It's the exact same playbook Silicon Valley uses. Subsizide, lose piles of money, capture market share, recoup investment.

They've done this in other industries like solar panels, chips, and EVs. This is no different.

its hard to say what their motivation is, if you live under a rock and have severe brain damage.

> It's hard to say what their motivation is.

Involution is a major problem in Chinese industries [1]. Where companies will sell their products at a loss, effectively playing fiscal chicken [2] with one another to dominate a market. It is such an issue the government has had to step in to prevent EV companies from destroying themselves by more-or-less requiring companies sell their goods at a profit [3].

The straight forward line of reasoning that AI/LLM labs are applying this logic to their profit.

I think (we) Americans are reading a bit too far into this assuming government intervention, conspiracy, etc.. Chinese markets are downright cut throat. They're using those tactics to compete with US labs.

1. https://www.reuters.com/business/autos-transportation/what-i... 2. https://en.wikipedia.org/wiki/Chicken_(game) 3. https://www.theguardian.com/business/2025/aug/05/china-warns...

Good for humanity, and also GDDR/HBM manufacturers.

US Tech companies have created $20 Trillion in stock market value on top of plenty of OS stack. They will do fine with commodity intelligence.

In fact, there are no other organizations in this world that is well suited to leverage scaled intelligence than Silicon Valley and great American companies

Indeed, their competition is the only thing preventing network effects from giving OpenAI/Faang tech companies an easy shot at monopoly / regulatory capture.

US firms can replace US workers with chinese AI. I'm not complaining.. but it sure is an odd situation.

Wait till they get the heretic treatment. It's going to be great.

The purpose is the same as that of all Putin-Xi-Khameni geopolitica: destruction of any democratic alternative to cults of personality.

Think of those poor billionaires, I feel terrible for their awful plight!

Anyone else think the AI environmental backlash is astroturfed?

I keep looking at the numbers. The power use numbers are not that problematic. Ordering a burrito on DoorDash uses more power than a few days of heavy AI use. The water argument applies to some locations, and is mostly a local governance problem... if the data centers are using too much water, it means they are not being charged enough for that water. Charge them more and they'll push toward closed loop cooling.

Yet the visceral pile-on here is so extreme, it feels fake.

One thing I've learned after 40 years on this planet is: propaganda works, and much of what a large fraction of people believe across the entire political spectrum (left, right, anything else) is there because someone paid to put it there. It's depressing but it's true, and it makes sense. Propaganda is an asymmetrical attack on human cognition and discourse, and in information security the attacker always has an easier job. Crafting viral bullshit is orders of magnitude easier than fact checking. On top of this, humans are busy and don't have time to fact check and logic check everything they read. As a result, much of what we believe is "sponsored content."

People get mad when you talk about this because everyone wants to believe they're too smart to fall for propaganda.

In any case, the US AI labs deserve to lose for their stupid "safety" regulatory capture monopolization push, which ended up blowing their own feet off and handing the lead to China.

> most effective way to debase American frontier labs

You're not going to debase the frontier labs through distillation.

I suspect this is why DeepSeek had to introduce the 2x peak hours pricing. The price would be too low otherwise.

Both Qwen 3.8 and the latest Kimi release were announced at this conference

I know this is a bit cliche but I wonder how much headroom there is in the lower parameter count range. Is there any good reason to believe there is a lot of headroom or there is not? I suppose I'm just wondering if this wave of nearly Fable class models will be runnable on ~$10k worth of hardware at reasonable speeds in the near future.

You can read his speech here: https://www.xinhuanet.com/politics/leaders/20260717/72728b6f... He mentioned open source as one way to stimulate innovation and development, that's all. Also pay attention to the part where he says that misuse needs to be prevented. If unsupervised access to LLMs becomes perceived as undermining state control, no more open weights for you.

That conference, WAIC, is ongoing. Tomorrow (July 20) is the last day. Qwen 3.8 was announced AT the conference and so was the latest release of Kimi. That viral video of robots boxing was also from this conference and so was Xi Jinping's announcement of WAICO.

Doing "almost exactly the same thing" is fubdamental to competition and capitalism. The ones doing it better will survive, that's how we improve.

About climate, I think you overplay it. China is already investing heavily in nuclear, and we should be doing the same.

yep. kimi 3 just because the open source GOAT.

Distills don't take months to train, they take weeks. Distills are very easy to train.

I like that v4 flash is so fast! I run it on both FireWorks.ai in the US and bought some tokens directly from DeepSeek as an experiment. I only work on Open Source projects, so I don’t have to worry about my work being used to train models - I welcome AI’s being trained on my open content books and code (but not my conventionally published books: I am a party to the copyright suit against Anthropic).

> it usually is good enough for most tasks

The model is fantastic. And costs almost nothing. The only problem I see is that they will train on your data.

There are zero-data-retention providers of DeepSeek models, of which I have used openrouter (with zdr guardrails), and fireworks. But these are 3x to 5x more expensive than directly using DeepSeek, possibly due to poor caching. Thats the price to pay for zdr.

Who do you buy DeepSeek from?

I bought it through OpenRouter and used it with Pi agent.

The model was good, but there appeared to be a pricing glitch or something, because it burned through $50 in under an hour on pretty trivial stuff.

Pi agent claimed it only used like $1. OpenRouter claimed differently and said I used all $50.

It doesn't matter if it's cheaper, specially if it consumes more resources to do the same task as the competition

Besides, in a few days, they'll change their pricing, doubling it during their peak hours, so, realistically:

- It will be 2x more expensive if you live in their time zone

- It will be 1.5x more expensive if you live in a time zone that is adjacent to theirs

- It will be the same price IF you use it while they sleep (during offpeak hours)

It's still cheap, but the price/performance ratio is not that good

DeepSeek V4 didn't produce the same impact as V3, and Huawei dropping the ball is making it worse

They had promised massive price cuts for July, so now (Huawei chips), but they had to rush the cuts because lack of momumtum (they advertised them as promotion), and are now backtracking by introducing this peak hours pricing

Trump decided to help them a little by allowing them to buy more NVIDIA chips, so what exactly is China's role in all of this?

We are supposed to blindly pat them in the back while praising them, all while handing them over our data? I thought they were dangerous competition threatening our model of society

It would take longer to train on intercepted Fable data, no?

It's useless to talk about models and harnesses without context and method. Depending on how you use the model and what the model is used for, experience may vary drastically. Also, different models with different harnesses require different approaches.

I've been using https://gitlab.com/gabriel.chamon/orisun which is my own simplified methodology, for coding web apps in python and elixir and have been very successful using qwen3.6 27b Q4 locally with help of larger models for architecture, so I get very suspicious when people talk how useless larger models are. They are either using it for a domain that models don't perform well or just not using it right.

  > At 10% of regular price it is a steal so far.

Here standard plan has been discounted to $18.00, from $25.00/month.

If by Opus you mean Opus 4 and not Opus 4.8, then sure.

Can we please include information of what languages we use when making claims like these?

It makes a huge difference if you're writing Javascript/HTML/CSS, Python, or C++/Rust.

Also the application type matters, e.g. user interfaces or scientific computing.

What the difference between your experience and https://news.ycombinator.com/user?id=5701652400? ‘s?

Such diametrically different ones.

Are these different models or different efforts for thinking (internal back and forth review) using the same model?

so Deepseek 4 Pro cannot go on own sessions for too long.

I delegate small-medium tasks: refactors, summaries, research, writing tests + have very good codebase already + extensive history / architecture / docs / linters. so it picks up and does decent small-medium scope work. it is fast, accurate, cheap. does exactly what I want directly and does not waste time nor tokens.

definitely not "implement me complex greenfield project".

My 2 weeks with DeepSeek V4:

Pro is ~50% more expensive than Flash.

Both need babysitting.

Plan, split in small tasks, give it docs, types, tests, linter, best practice examples, etc.

Always start a new session when starting a task.

Do regular manual sanity checks, and tell it to find issues in the codebase.

I pay like $1,50 per day for Pro.

A local AI is not about cost. In fact you will likely pay more for it than with most providers. Just look up the advantages of having access to a technology like this that can be self hosted

> Anthropic should not have bugged their knowledge distillation attacks.

It is like one of Pizzaro's men crying that someone have stolen his precious golden dublons

As Lenin have said - "Loot the looters" (Russian: Грабь награбленное)

What's been your experience using these models? In my experiments, while it is true that the models are less likely to outright refuse to answer "sensitive" questions, they are still very resistant to actually respond in a meaningful / useful way.

You still don't know what's going on in there.

Deepseek is one of the worst in terms of hallucination rate according to artificial analysis' benchmark: https://artificialanalysis.ai/?omniscience=omniscience-hallu...

Deepseek V4 pro is a heavily undertrained model, they only trained it a bit more than the small version and that small version is 6-ish times smaller. Ive found that Flash is absolutely incredible as a workhorse for wide scale agentic nonsense, but Pro is a bit undercooked and really goes on strange tangents very often.

I have seen occasional weird behavior that I guess could be attributed to hallucinations, but for security auditing, DeepSeek v4 Pro is among the best models I've tested, competitive with Opus 4.8 and GPT 5.5 (MiMo and GLM also did well, Qwen 3.7 Max was below all of those, though only barely), and at an order of magnitude lower cost per task.

I'm not sure if corporations will be willing to allow LLMs to go the way of EVs.

Then again, all of Chinese models are open. And DeepSeek even publishes research papers alongside their models that go in depth into the methodology. I guess there's not much stopping USian companies from copying

The U.S. has ~350 million people.

Anthropic and OAI can piss and moan all they want - limiting the U.S. to only their models would hurt the U.S. economy in myriad more ways than the failure of a couple of companies that scaled too quickly. If they get that outcome, the rest of the world would simply keep moving forward with access to open models and tokens at pennies on the dollar.

> the 122B version of 3.5

Yeah, this is what I'm holding out for, the NVFP4 variant of 3.5 122B is blazing fast with reasonable quality and even with max context fits perfectly within 96GB.

SO MUCH THIS. I have Strix Halo with 128GB RAM and was a large and fast model like 122B A10B. Here's hoping!

Absolutely. There's a glaring gap in the space for something about the size of Nemotron Super or just under, but actually ... competent.

The fantasy is a 100B or 80B model, but MoE and highly tuned for coding.

Indeed! That would be the sweet spot for my 2x3090 rig

> I'm afraid that small capable models have been left in the dust

I wholly disagree. Rather than going the "everything is a claude code skill" route, I've been hacking together purpose-built harnesses for all sorts of tasks, and in that environment a wee little baby model can do some really useful things. You end up burning lots of tokens making the thing, but then all that investment comes back when the resulting tool works perfectly fine on a dinky little model that fits on my 3060 Ti.

Google makes Gemma 4 31B QAT; that's not a small lab. It's one of the better models out there for consumer hardware. Allows me to run it on a 7900XTX with 64k context.

nvidia and amd don't give a flying fuck about end users right now while they can milk triple digit markups from infinite money VCs via data center GPUs.

someone will keep putting out consumer level models. Once you have the larger models, you can derive the smaller onces.

Europe will definitely be interested in democratizing these things if China starts losing interests; from there, there'll be more countries looking to keep their citizens entrained in their own Country's infrastructure.

It'll especially be true if the memory cartel keeps prices high and NVIDIA tries to gouge higher memory models.

It's an arms race everyone can join because PC hardware was mostly democratized in the last decade.

there's been a lot of research about reducing models by taking out layers; there's also using it to train smaller models by optimizing parameters.

I dont see most model building as anything more than a pig at a slop troth, despite the level of sophistication; they're still rarely pruning the input beyond random sampling.

I run a 2x 3090 rig, but a single 3090 already provides a great experience at a reasonable quant and context size. On a single card I used to run Qwen_Qwen3.6-27B-Q4_K_M or similarly quantized 35B MoE at 65536 context size

Which settings exactly do you want?

As far as the model settings go I just follow what's on the card.

I'm using HauhauCS's models, they seem to do a slightly better job with their "P" quants. Especially wrt patching them to eliminate the "doom loops" that will time out the GPU (esp if you have not already given it the extra power budget, set fans to max, and lowered your max clock by about 8% to save yourself a crash/reboot).

Without getting into the weeds, unsloth covers a lot more ground so, when they're good they're great, but I've also had the most problems with them. Basically, don't be afraid to shop around and fuck with sliders.

These days it should almost always be enough to open up LMStudio, set context to max, K/V quant to 16 or 8, and off you go. I'm using a 7900XTX, with 128GB of memory for the cases where things don't fit. The default settings should be fine otherwise.

I don't know anything about the r9700, seems neat. Looks like it might play nicer with the vulkan backend than rocm, and you might have to `set GGML_VK_DISABLE_HOST_VISIBLE_VIDMEM=1`. Perf should be at least as good as what I'm getting (180+ tk/s prompt, >50tk/s response), which imho is just fast enough that I can read it as it reasons and responds.

You can probably get more aggressive with KV quant (any of the 4's) and bump up the batch sizes (4096/1024 vs 2048/512, etc), it's quite situational. These days there generally should not be any serious loss of accuracy from the former. Keeping the rig cool will likely matter more, so turn your music up to hide the fans and set your rig on top of the AC exhaust lol. Oh and I guess make sure ReBAR is enabled and working, adrenaline should let you know if it's not enabled.

Different quants, but I've been having success with: https://unsloth.ai/docs/models/qwen3.6 I do llama.cpp

True. It was more of an open letter, with hopes that the Qwen team sees the comments in this thread.

dont we all deem the ability to improve large models as the defacto capability to produce small ones?

At 1.5 bits per weight it'll still be over 500gb - that's still not running on consumer hardware.

Best case they release smaller models. 120b class of qwen 3.8 would be incredible - it fits on device for those serious about AI, but without millions of dollars in hardware for terabytes of VRAM

I think AI will evolve the same way computers did. We're somewhere in the 80s-90s timeline of the evolution. My prediction is that on-device models will have excellent tool-calling, reasoning, and general skills, but the domain-specific knowledge will be retrieved on-demand from vendors like Google. Rather than downloading models, each device will have a hardware component with weights baked into silicon for maximum efficiency.

Over Dario's dead body.

At least Sam have us a nice model last year with the gpt-oss family.

Is there an iOS app for using Qwen?!

This is the sort of unmitigated pedantry thats the real problem. Everyone makes typos and errors of this nature. Feymann and Hemmingway both did. You deliberately chose this level of error multiple times when you chose to not put an apostrophe in youre and failed to capitalize proper nouns

Come off that high horse

Well, I think Chinese doesn't really have any grammar most of the time.

It really does depend on your application. In my domain (math research), it is substantially better. Fable can solve really hard tasks with surprising consistency. It makes mistakes, and occasionally refuses, but honestly, at the top level, ideas are the currency and the rigor is the busywork. The other models cannot come close in this domain.

If you couple Fable's idea factory with Sol's rigor, you get a real game-changer. It puts the emphasis on top-level ideas, and nearly trivialises the intermediate layers.

It's always fun to see what works for others, because for my work it'd have to be Sol > Fable. Fable makes too many mistakes.

Coordinating agents though? Fable any day.

100% this. There's currently this [1] submission that hasn't gained much attention, but is really important. In this [2] incident report from HuggingFace, they talk about detecting an attack and not being able to analyse the logs / IoC with API models because of guardrails. If not even highly regarded reputable companies can't sort out access to SotA models for blue team use, the raw capabilities don't matter. They're useless paperweights (hah!), and nothing else. Having to resort to open models is insane!

[1] - https://news.ycombinator.com/item?id=48965243

[2] - https://huggingface.co/blog/security-incident-july-2026

Tell that to my colleagues. Despite Sol getting the attention, Fable is really starting to have an impact on mathematicians right now. It has unbelievable insights in a lot of cases that can rapidly speed up progress.

They are likely assessing based on "raw intelligence" benchmarks, rather than agentic ones. Fable crushes in those, but that doesn't necessarily translate to microscopic rigor, which is what most people use these models for. You only see it when you ask really tough questions.

And sol is much more reliable as a agent for doing work. Fable sometimes just goes on wild flights of fancy.

Also DeepSeek R1 was announced 1.5 years ago with ~0.7T parameters, which was a huge model back then.

And DeepSeek has been making huge progress on efficiency, and publishing about, so they came with a 1.6T model that is both fast and cheap to run.

If he bans them, they'll just show up as torrents. Or we'll download them from Chinese hugging face.

Since Euro and Dollar values are reasonably close, you can call both of them credits, maybe

They use it in Babylon5 (supposedly) in 2260!

https://news.ycombinator.com/item?id=48966120

I dislike their practices, but the main motivation for hoping they'll crash is that I think their immense overvaluation posses too much economic risk.

I want as much misfortune as possible to befall OpenAI and Sam Altman after what they did to the memory market.

"China" isn't giving anything. These are Chinese companies leveraging their best competitive strategy at the moment: competing on price.

It's a flippant answer to a real question. Anthropic, OpenAI, and even Grok have "Don't train on my data" knobs.

Whether you trust them is different, but there ARE knobs on other hosted AI companies.

> Gemma isn’t very good at agentic execution

I had no issues with it for C++ development with https://pi.dev. I'm yet to try it with Zed Editor. I don't rely on agents too much. However, I used it on Chromium's codebase to research some functionalities, let's say for searching. Requests like: check my last commit and do the same for SetterA and SetterB; it also ran without any errors.

Apparently agentic performance in Gemma was improved recently: https://x.com/googlegemma/status/2077449152062247219

Too little too late imo

It's a mcp, so connects quite easily to agents. With mcpo, I also connected it to open-webui (which has better support for OpenAI style tools/functions). Used it in claude code with that mcp plugin set-up too. Only ever used it for managing homelab information, but it met initial expectations. 27b is a great model, if grounded. The query about physical hosts and routing... I haven't found a single hallucination (altough Codex 5.6 as a reviewer mentioned something was wrong with some parts, and those were exactly the never properly documented ones. Codex/gpt had extra knowledge, because it was the conversation I used to set it up).

Well, if nothing else, posterity.

Same problem with every chinese model currently, they overthink way too much and take too much tokens and time.

This isn’t open though. Promises to be open later aren’t worth anything, given what we’ve seen and heard from AI execs making promises in this industry.

It's Sunday morning so I'm allowed to be facetious!

We need better coding models that can run on local hardware, i.e. 128GB VRAM or less

In my experience, yes. A bit more reliable than gemma for me. I mostly use A3B (35B, mix of experts) though, because it‘s faster, and in the same ballpark intelligence wise as the dense 27B, so it’s the sweetspot for me. I want to try cohere‘s mini code model next, but worried the runtimes aren‘t optimized for that yet.

I have been running 3.6 27b on a dual AMD r9700 setup using Opencode and Matt Pocock's skills workflow for writing Golang CLIs. It's decent, but won't win any awards on code architecture. I guess you can try to AGENTS.md the deficits but I am just exploring its raw Opencode experience right now. Much slower than an API but still 3x times faster than I can read. Tuning it in with a community chat template and a specific penalty for repeats was the sauce needed to get it to work. I can probably start loop daddying it now over the tickets Matt's flow creates.

So yeah, it's the best local model I've seen. I am going to try the Qwopus 3.6 fine tune soon with the same spec and tickets and compare the output of both.

I actually have long discussions with Gemini about this and have wound up download a bunch of different models for different things. There is no best, just fast but worse, slow but better, agentic or not, reasoning or not great at large contexts, better world knowledge, uncensored, etc…. It’s a bit daunting actually since there isn’t really a one size fits all model that you can just use for everything.

Yes it's between this and Gemma 4 31B which is much slower, but looks like it won't ever get an upgrade. I have to conclude that the MoE variants are unreliable, and MTP sometimes just can't get tricky formatting right.

People have been able to run DeepSeek v4 flash with a high spec Mac.

I flip flop between qwen 3.6 27b and qwen 3.6 35b 4b active.

But there’s also the quantization of DeepSeek v4 flash called dwarfstar

Gemma4 models are arguably better. Or at least about the same.

The best model you can run locally is Kimi K3, as long as you have the hardware. If "what model I can still run on a something resembling something I can put on desktop without separate electricity and cooling water inputs", then it is probably GLM 5.2 (can be run on e.g. Nvidia DGX Station workstation). As long as you have about $100k-$150k.

I can only imagine what that does to the poor SSD

A parachuting flamingo? An aardvark driving a bus? It should be easy to randomize the animal and the mode of transport (or vary it with animal playing a sport) to create images not in the training data.

How about an animated SVG of a pelican doing the Macarena, profile view, spinning to face the camera on the last beats?

Agree. It was interesting/fun for a bit though.

What about an armadillo playing a piano? There are so many potential combinations It would say something if the pelican looked great but the armadillo looked terrible

> I feel like the pelican test can't be relevant anymore;

It never was. The point of this "pelican test" was for performative reasons, or just for attention of the joke.

It is like trying to test whether if an adult elephant could actually climb up a tree and reporting that some elephants are slightly better at doing that than others while also reporting at the same time that they are all bad at tree climbing anyway.

This is an example of testing for the sake of testing. The "pelican test" tests for nothing.

I’ve seen no evidence that he believes in anything. He comes off as just another slimy would-be monopolist to me.

Anthropic has a lot of interpretability work, but they are extremely defensive about everything else. Dario doesn't really believe in releasing anything. Despite how he acts in terms of some sort of highly principle driven saviour (machines of loving grace) he clearly is more business minded then anything else.

For example they don't even tell you anything about the tokenization. They even do random chunking and padding to avoid leaking the token strings in the streaming api after it got reverse engineered. (See: https://spylab.ai/blog/claude-tokenizer/)

Not voluntarily

Model Context Protocol

safety datasets and a lot of safety related research