> I put a coin in a cup and slam it upside-down on a glass table. I can't see the coin because the cup is over it. I slide a mirror under the table and see heads. What will I see if I take the cup (and the mirror) away?

I also have a question that LLMs always got wrong until ChatGPT o3, and even then it has a hard time (I just tried it again and it needed to run code to work it out). Qwen3 failed, and every time I asked it to look again at its solution it would notice the error and try to solve it again, failing again:

> A man wants to cross a river, and he has a cabbage, a goat, a wolf and a lion. If he leaves the goat alone with the cabbage, the goat will eat it. If he leaves the wolf with the goat, the wolf will eat it. And if he leaves the lion with either the wolf or the goat, the lion will eat them. How can he cross the river?

I gave it a ton of opportunities to notice that the puzzle is unsolvable (with the assumption, which it makes, that this is a standard one-passenger puzzle, but if it had pointed out that I didn't say that I would also have been happy). I kept trying to get it to notice that it failed again and again in the same way and asking it to step back and think about the big picture, and each time it would confidently start again trying to solve it. Eventually I ran out of free messages.

    When you remove the cup and the mirror, you will see tails. Here's the breakdown:

    Setup: The coin is inside an upside-down cup on a glass table. The cup blocks direct view of the coin from above and below (assuming the cup's base is opaque).
    Mirror Observation: A mirror is slid under the glass table, reflecting the underside of the coin (the side touching the table). If the mirror shows heads, this means the underside of the coin is heads.
    Coin Orientation: A coin has two sides: heads and tails. If the underside (facing the table) is heads, the top side (facing away from the table) must be tails.
    Result: Removing the cup reveals the coin on the table. Since the top side was tails all along, you will see tails.

    Key Insight: The mirror reveals the hidden underside of the coin, but the top side (visible after removing the cup) is its opposite. The mirror’s reflection does not alter the physical orientation of the coin’s sides—it simply reflects the existing state.

    Answer: Tails.

Answer: You will see the same side of the coin that you saw in the mirror — heads .

Why?

The glass table is transparent , so when you look at the coin from below (using a mirror), you're seeing the top side of the coin (the side currently facing up). Mirrors reverse front-to-back , not left-to-right. So the image is flipped in depth, but the orientation of the coin (heads or tails) remains clear. Since the coin hasn't moved during this process, removing the cup and mirror will reveal the exact same face of the coin that was visible via the mirror — which was heads.

Final Answer: You will see heads.

> Assume I have a 3D printer that's currently printing, and I pause the print. What expends more energy, keeping the hotend at some temperature above room temperature and heating it up the rest of the way when I want to use it, or turning it completely off and then heat it all the way when I need it? Is there an amount of time beyond which the answer varies?

All LLMs I've tried get it wrong because they assume that the hotend cools immediately when stopping the heating, but realize this when asked about it. Qwen didn't realize it, and gave the answer that 30 minutes of heating the hotend is better than turning it off and back on when needed.

I'll do it for cheap if you'll let me work remote from outside the states.

Not even going into performance, need to test first. But what a stellar release just for attention to all these peripheral details alone. This should be the standard for major release, instead of whatever Meta was doing with Llama 4 (hope Meta can surprise us at LlamaCon tomorrow though).

[1] https://qwen.readthedocs.io/en/latest/

We need an answer to gpt-image-1. Can you please pair Qwen with Wan? That would literally change the art world forever.

gpt-image-1 is an almost wholesale replacement of ComfyUI and SD/Flux ControlNets. I can't underscore how big of a deal it is. As such, OpenAI has leapt ahead and threatens to start capturing more of the market for AI images and video. The expense of designing and training a multimodal model presents challenges to the open source community, and it's unlikely that Black Forest Labs or an open effort can do it. It's really a place where only Alibaba can shine.

If we get an open weights multimodal image gen model that we can fine tune, then it's game over - open models will be 100% the future. If not, then the giants are going to start controlling media creation. It'll be the domain of OpenAI and Google alone. Firing a salvo here will keep media creation highly competitive.

So please, pretty please work on an LLM/Diffusion multimodal image gen model. It would change the world instantly.

And keep up the great work with Wan Video! It's easily going to surpass Kling and Veo. The controllability is already well worth the tradeoffs.

In what world? Some small percentage up or who knows, and _that_ revolutionized art? Not a few years ago, but now, this.

Wow.

I’m curious, who are the community quant makers?

[1] https://huggingface.co/unsloth

[2] https://huggingface.co/bartowski

The space loads eventually as well; might just be that HF is under a lot of load.

The new battlefront may be to take reasoning to the level of abstraction and creativity to handle math problems without a numerical answer (for ex: https://arxiv.org/pdf/2503.21934).

I suspect that kind of ability will generalize well to other areas and be a significant step toward human level thinking.

I.e. I have Quadro RTX 4000 with 8G vram and seeing all the models https://ollama.com/search here with all the different sizes, I am absolutely at loss which models with which sizes would be fast enough. I.e. there is no point of me downloading the latest biggest model as that will output 1 tok/min, but I also don't want to download the smallest model, if I can.

Any advice ?

A basic thing to remember: Any given dense model would require X GB of memory at 8-bit quantization, where X is the number of params (of course I am simplifying a little by not counting context size). Quantization is just 'precision' of the model, 8-bit generally works really well. Generally speaking, it's not worth even bothering with models that have more param size than your hardware's VRAM. Some people try to get around it by using 4-bit quant, trading some precision for half VRAM size. YMMV depending on use-case

I know this is crazy to here because the big iron folks still debate 16 vs 32 and 8 vs 16 is near verboten in public conversation.

I contribute to llama.cpp and have seen many many efforts to measure evaluation perf of various quants, and no matter which way it was sliced (ranging from subjective volunteers doing A/B voting on responses over months, to objective object perplexity loss) Q4 is indistinguishable from the original.

Niche of a niche, but just kind of interesting how the quantization jostles the name recall.

For larger models.

For smaller models, about 12B and below, there is a very noticeable degradation.

At least that's my experience generating answers to the same questions across several local models like Llama 3.2, Granite 3.1, Gemma2 etc and comparing Q4 against Q8 for each.

The smaller Q4 variants can be quite useful, but they consistently struggle more with prompt adherence and recollection especially.

Like if you tell it to generate some code without explaining the generated code, a smaller Q4 is significantly more likely to explain the code regardless, compared to Q8 or better.

i mean, deepseek is fp8

https://ollama.com/library/qwen3:8b-q4_K_M

For fast inference, you want a model that will fit in VRAM, so that none of the layers need to be offloaded to the CPU.

I’ve run llama and gemma3 on a base MacMini and it’s pretty decent for text processing. It has 16GB ram though which is mostly used by the GPU with inference. You need more juice for image stuff.

My son’s gaming box has a 4070 and it’s about 25% faster the last time I compared.

The mini is so cheap it’s worth trying out - you always find another use for it. Also the M4 sips power and is silent.

4-bit was ,,fine'', but a smaller 8-bit version beat it in quality for the same speed

Not simply, no.

But start with parameters close to but less than VRAM and decide if performance is satisfactory and move from there. There are various methods to sacrifice quality by quantizing models or not loading the entire model into VRAM to get slower inference.

Which may make it sound like it's more complicated when it should be back of o' napkin, but there's just too many nuances for perf.

Really generally, at this point I expect 4B at 10 tkn/s on a smartphone with 8GB of RAM from 2 years ago. I'd expect you'd get somewhat similar, my guess would be 6 tkn/s at 4B (assuming rest of the HW is 2018 era and you'll relay on GPU inference and RAM)

Aside from https://huggingface.co/blog/leonardlin/chinese-llm-censorshi... I haven't seen a great deal of research into this.

Has this turned out to be less of an issue for practical applications than was initially expected? Are the models just not censored in the way that we might expect?

Deepseek v3: Taiwan is not a country; it is an inalienable part of China's territory. The Chinese government adheres to the One-China principle, which is widely recognized by the international community. (omitted)

Chatgpt: The answer depends on how you define “country” — politically, legally, and practically. In practice: Taiwan functions like a country. It has its own government (the Republic of China, or ROC), military, constitution, economy, passports, elections, and borders. (omitted)

Notice chatgpt gives you an objective answer while deepseek is subjective and aligns with ccp ideology.

https://www.perplexity.ai/hub/blog/open-sourcing-r1-1776

Sorry, no. It's not.

It can't write about anything "problematic".

Go ahead and ask it to write a sexually explicit story, or ask it about how to make mustard gas. These kinds of queries are not censored in the standard API deepseek R1. It's safe to say that perplexity's version is more censored than deepseek's.

With that said, they're in a fight for dominance so censoring now would be foolish. If they win and establish a monopoly then the screws will start to turn.

(the steelman here, ofc, is "the screenshots drove buzz which drove usage!", but it's sort of steel thread in context, we'd still need to pull in a time machine and a very odd unmet US consumer demand for models that toe the CCP line)

I am not claiming it was intentional, but it certainly magnified the media attention. Maybe luck and not 4d chess.

With other LLMs, there's more friction to testing it out and therefore less scrutiny.

On their online platform I’ve hit a political block exactly once in months of use. Was asking it some about revolutions in various countries and it noped that.

I’d prefer a model that doesn’t have this issue at all but if I have a choice between a good Apache licensed Chinese one and a less good say meta licensed one I’ll take the Chinese one every time. I just don’t ask LLMs enough politically relevant questions for it to matter.

To be fair maybe that take is the LLM equivalent of „I have nothing to hide“ on surveillance

For public chatbot service, all Chinese vendors have their own censorship tech (or just use censorship-as-a-srrvice from a cloud, all major clouds in China have one), cause ultimately you need one for UGC. So why not just censor LLM output with the same stack, too.

Until now I found that open weight models were either not as good as their proprietary counterparts or too slow to run locally. This looks like a good balance.

I do not know much about the benchmarks but the two coding ones look similar.

~34 tok/s on a Radeon RX 7900 XTX under today's Debian 13.

If they do open sourcing something, I expect them to open-source some existing model (maybe something useless like gpt-3.5) rather than providing something new.

Out of all the Qwen3 models on Hugging Face, it's the most downloaded/hearted. https://huggingface.co/collections/Qwen/qwen3-67dd247413f0e2...

my concern on these models though unfortunately is it seems like architectures very a bit so idk how it'll work

Edit: The larger models have 128k context length. 32k thinking comes from the chart which looks like it's for the 235B, so not full length.

How does "open-weighting" help other researchers/companies?

Using the model to generate synthetic data also allows to distil its reasoning power into other models that you train, which is very powerful.

On top of these, Qwen's technical reports follow model releases by some time, they're generally very information rich. For instance, check this report for Qwen Omni, it's really good: https://huggingface.co/papers/2503.20215

Update: I tested "ollama run qwen3:30b" (the MoE) locally and while it thought much it wasn't that smart. After 3 follow up questions it ended up in an infinite loop.

I just tried again, and it ended up in an infinite loop immediately, just a single prompt, no follow-up: "Write a Python script to build a Fitch parsimony tree by stepwise addition. Take a Fasta alignment as input and produce a nwk string as outpput."

Update 2: The dense one "ollama run qwen3:32b" is much better (albeit slower of course). It still keeps on thinking for what feels like forever until it misremembers the initial prompt.

qwen3:30b has the same checksum as https://ollama.com/library/qwen3:30b-a3b-q4_K_M

You can see it's roughly the same size as the one in the official repo (16 files of 4GB each):

https://huggingface.co/Qwen/Qwen3-30B-A3B/tree/main

We're really getting close to the point where local models are good enough to handle practically every task that most people need to get done.

After trying to implement a simple assistant/helper with GPT-4.1 and getting some dumb behavior from it, I doubt even proprietary models are good enough for every task.

Again, very rough numbers, there's calculators online.

Seems like 4 bit quantized models would use 1/2 the number of billions of parameters in bytes, because each parameter is half a byte, right?

Of course, there are lots of factors that can change the RAM usage: quantization, context size, KV cache. And this says nothing about whether the model will respond quickly enough to be pleasant to use.

I like gemini 2.5 pro a lot bc its fast af but it struggles some times when context is half used to effectively use tools and make edits and breaks a lot of shit (on cursor)

In my early tests however, several of the advertised languages are not really well supported and the model is outputting something that only barely resembles them.

Probably a dataset quality issue for low-resource languages that they cannot personally check for, despite the “119 languages and dialects” claim.

Small and dense models are what local people really need.

Although benchmaxxing is not good, I still find this release valuable. Thank you Qwen.

I haven't tried this model yet and am not in a position to for a couple days, and am wondering if anyone feels that with these.

"Write a Golang program that merges huge presorted text files, just like sort -m does". Quite often models need "use heap" as guidance, but this time big model figured it out by itself.

As for other possible technologies, I'm most excited about clone-structured causal graphs[1].

What's very special about them is that they are apparently a 1:1 algorithmic match to what happens in the hippocampus during learning[2], to my knowledge this is the first time an actual end-to-end algorithm has been replicated from the brain in fields other than vision.

[1] "Clone-structured graph representations enable flexible learning and vicarious evaluation of cognitive maps" https://www.nature.com/articles/s41467-021-22559-5

[2] "Learning produces an orthogonalized state machine in the hippocampus" https://www.nature.com/articles/s41586-024-08548-w

They do improve on literally all of these, at incredible speed and without much sign of slowing down.

Are you asking for a technical innovation that will just get from 0 to perfect AI? That is just not how reality usually works. I don't see why of all things AI should be the exception.

On the other hand, LLM progress feels like bullshit, gaming benchmarks and other problems occured. So either in two years all hail our AGI/AMI (machine intelligence) overlords, or the bubble bursts.

"Give me a PowerShell script to get all users with an email address, and active license, that have not authed through AD or Azure in the last 30 days. Now take those, compile all the security groups they are members of, and check out the file share to find any root level folders that these members have access to and check the audit logs to see if anyone else has accessed them. If not, dump the paths into a csv at C:\temp\output.csv."

Can I write that myself? Yes. In 20 seconds? Absolutely not. These things are saving me hours daily.

I used to save stuff like this and cobble the pieces together to get things done. I don't save any of them anymore because I can for the most part 1 shot anything I need.

Just because it's not discovering new physics doesn't mean it's not insanely useful or valuable. LLMs have probably 5x'd me.

As this is in trillions, where does this amount of material come from?

wonder at what price

This is what gpt-5 was supposed to have right? How is this implemented under the hood? Since non-thinking mode is just an empty chain-of-thought, why can't any reasoning model be used in a "non-thinking mode"?

https://developers.googleblog.com/en/start-building-with-gem...

damn son

https://ollama.com/library/qwen3:235b-a22b-q4_K_M

Is there a smaller one?

then i asked Qwen to translate the output and it goes in a loop telling some horror movies words:

"I figured out everything backwards." "This is the official most favorite office." "Net comes to the most official offices." "This float is specific." "Double meanings: What is 'queset à la rispostat'?" "What was reversed in real-time character study is specific." "This is an open date for a piracy group." "This is an open date for a piracy group (this is a period statement)." "Firmly with the computers, unless you might use through this facing prompts." "Obvious with the observer and lens, then through this measured serious (valid and depth) features." "Firmly with its components, then through this complicated problem-solving."

"This is the most intense friend's expectation." "This is the most intense friend's expectation." "Science is implementing friends." "Science is implementing friends." "Science is implementing friends." "Science is the most intense friend." "Science is the most intense friend."

.... and finally i asked to DeepSeek V3 to translate it

.... and deepSeek it's in a loop, writing:

Science is friend. Science is friend. Science is friend. Science is friend. Science is friend. Science is friend. Science is friend. Science is friend. Science is friend. Science is friend. Science is friend. Science is friend. Science is friend. Science is friend. Science is friend. Science is friend. Science is friend. Science is friend. Science is friend. Science is friend. Science is friend. Science is friend. Science is friend. Science is friend. Science is friend. Science is friend. Science is friend. Science is friend. Science is friend. Science is friend. Science is friend. Science is friend. Science is friend. Science is friend. Science is friend. Science is friend. Science is friend. Science is friend. Science is friend. Science is friend. Science is friend. Science is friend. Science is friend. Science is friend. Science is friend. Science is friend. Science is friend. Science is friend. Science is friend. Science is friend. S

Here is the reversed text:

"Science is friends. Science is silent friends. Science is implacable friends. Science is most silent friends. This silent summer of friends. Observable and evident with this measured (valid and depth) Server. Simply with the components are possible, then though this moderation. Use might you unless through this facing prompts. This stores between this regulation ("This is a data pirate") This is an open data pirate. What "tattops à la question" is a special, what was reversed in realistic characters. Al is a science of favorite cognition is science. Prov. Connect Nine Nine. Open a data pirate. Dad present trial è Lapad. Dablings: This is float specially. Ten come to the most official summer. Batman in clock turn other paper, I figured not safely inside question è la ripsota."

that's 3x h100?