I'm still sad that we haven't seen a new Taalas style chip a la https://chatjimmy.ai/. Smaller models are good enough now to make that insane burst of tokens so useful.
I don't know the model behind this, but it is absurdly bad.
> Write me a coherent paragraph in French, without ever using the letter "e".
> Voilà une phrase claire et concise : "Le village est situé dans les montagnes. Le soleil est haut. Il y a des animaux dans le village. Il pleut dans les montagnes."
I suppose this is just a demo of how fast an LLM can be, I wonder if there are tradeoffs with larger/smarter models. Also, for a human usage, at what point are tokens generated fast enough that it's pretty much instant? My bet is below 1000 tps
Why ask this when we know that LLMs are not good at the character level. They run on tokens, not characters. In fact, they don't even see the characters, unless you do special tricks.
I asked it to translate your sentence to English and it did fine. In less than a fraction of a second.
To be fair, you picked a well-known tricky benchmark for LLMs: When working on an embedding spelling disappears after the embedding level. I imagine modern frontier models have tools that let them read back their input to work around this issue.
That's the problem with etching a model onto a chip: by the time you've designed the chip, manufactured it, tested it, shipped it, and deployed it, the model will be hopelessly outdated (with the current improvement rates). And when you want to update, you have to buy new chips instead of just uploading a new model file like now. When Taalas announced their chip, the model was already 1.5 years old (stone age by current standards). It's their first chip, so maybe they can streamline it, but the problem of having to update hardware every few months to keep up with the industry is not going anywhere.
You can upload different weights and even do LoRAs. The chip architecture is interesting, the first (n) layers are the sane, so you can change architecture by adding (m) layers. Plausible that this is sufficiently flexible enough for several generations of real world applications. For example, we still use 45nm general purpose silicon for automotive, e.g.
What they did had never been done before. Now we see that it's possible, there are plenty of models to choose from that could be etched into silicon. In the next year or two, I think these smaller models might plateau, and there may be some on-device niche they can fill.
for comparison, Qwen3.8-Flash-Next only requires 6B parameters for computation, but stores 125B, 51B of those can be comfortably offloaded as they're not actively used in decode but a single token look up.
The Quant iQ4 of this model loads, then, in ~60GB of vram, and on disk it's 85GB.
So if you could etch it, you'd need a ~25GB ssd chip and 60GB of vram.
The vram costs likely contributed to these things being out of reach of the current economic cycle.
I'm not smart enough to know why; I do know that 27B is greater for short/interactive on blackwell, but the intellgence leap of the MoE in Qwen3.8-Flash-Next is quite remarkable.
I'm pretty convinced the pathway to local models will be MoE, especially if they can find a way to keep tweasing out things like PLE into the slow bandwidth lanes.
The Taalas architecture makes loading weights free, but they basically has to pay the same silicon for every weight, whether it's used or not. MoE models are more efficient than dense models per weight you load, but less efficient per weight you have to store.
MoE models are the path to local models with traditional system architectures, but they are antithetical to what Taalas was doing. If you spent all the money to etch 125B weights into silicon, you'd want to activate them all for each token, instead of only touching 6B. You cannot match the 125B sparse model with a 27B dense one, but you might be able to match it with a 60B or so one.
Luna gets the button wrong, and in the same way Grok/MiMo did. Looking into it more, it's because Luna actually searched my computer for similar builds, found the ones that I did for grok/mimo, and referenced their files. Astra is still the best by a significant margin in my eyes. Far more polish, better page transitions, effects that aren't overcooked and take into account the page. Better contrast.
Overall it follows image designs quite well, but it did ignore asks to animate page transitions. Additionally it's the least performant of the ones I've built with Astra/Grok/MiMo, despite using a lot of the same code. I'd rate it just below Astra in capability, but still solidly second place.
For comparison with other drops this week + current #1:
Great, I opened several of these in new tabs in Firefox and the entire browser froze and I had to kill it. Can't tell you which one caused it though. I have plenty of RAM also.
Interesting. I just opened all 3 on a fresh install of firefox with no issues - can I ask what OS / do you have any js disabled / do you have webgl disabled?
Worth noting though that GLM 5.3 isn't multi-modal, so it doesn't have a vision layer. It is quite clever and hacks around it pretty effectively however. I'm running a deepseek 4 build now and will reply shortly with that.
The gist of it though is I take a prompt, expand it into a json blob specifying structure/palette/positioning of elements/etc, feed that into a diffusion model to output a few choices. Once I lock in a choice I take the pixel output + json blob and use it as input into followup pages. The json helps preserve the brand across multiple pages.
Once I have all the inputs I take their corresponding image+json blobs and feed them into an agent to create a web implementation.
For image models, diffui currently uses gpt-image-2.5, mai-image-2.6, and very, very rarely a post-trained version of flux 2 dev I've made for web design, though that one will be deprecated soon.
Overall this felt like the weakest of the three. Ultraspeed was fast as far as tokens per second goes, but it overthought quite a lot of things resulting it in having one of the longest build times. That overthinking didn't lead to better results either - note the statue with the cropped off arm. It's also the worst implementation of the dynamic lighting effect / displacement effect - the background especially has some significant distortion. Astra was the only one that seemed to understand that displacement should happen less the further something is in the distance.
For me it seems like this model’s overthinking should be tamed.
Perhaps instead of giving it a one shot task with vague prompt. I wonder how it can perform with much more detailed and constrained prompt (can you please elaborate more on the level of detailness and ambiguity that the prompt is and where does this model seem to overthink the most?)
Also are there any ways to tame such overthinking of models in general?
I hope that once models start becoming smart enough (I think for me it’s already there) or becoming genuinely the Sota. They then start focusing a lot more on optimizing token usage
Additional prompt instructions: "Add scrolling clouds behind the statues. Dynamically light the statues based on mouse position. Use diffui to generate the normal maps/depth maps/roughness maps of the objects, and to separate out the assets on to different layers."
Overall I find these models are getting good at following image as a source of instructions, but their refinement of the output varies heavily between the models. Astra's final output feels more polished, has better visual contrast, and the animations between the pages are smoother. Grok also chose to light all of the background elements, which imo overcooks it a bit.
Still though, for the price it's a great starting point.
I run a prompt-to-ui design site that uses image models for the design process[1]. The text rendering especially makes this model deeply interesting to me, despite the license. Here are some tests using my harness comparing the outputs of gpt-image-2 and qwen 2.1:
The text rendering definitely is much, much better than anything else on the open weights market right now. Small text fidelity is quite good. It seems like the text encoder however gets a little bit overloaded with larger prompts - note the presence of hex codes in the design output, those were inputs from the expanded prompt.
I'll be trying a post-training run on this for web design, it has some serious potential.
My harness expands the prompt into a json representation that specifies layout much more rigorously, which is why you see such that amount of alignment between the two.
That internal json backing helps significantly when you want to maintain consistent design system components/patterns across multiple pages. The aligned layout is it working as intended.
Totally normal for modern models due to training on the same datasets supplied by third parties, dataset contamination, and mode collapse, especially for simple prompts that don't have enough semantic capacity. -isms are often very similar even without distillation, and tend to come and go in waves along with model generations.
Agreed. There's also a lot of bad tinging/yellow saturation that very much reminds me of early gpt-image outputs on a lot of the non-cherry picked stuff I've been seeing on Twitter/Reddit.
A lot of people were putting ZiT as a refiner downstream in early Qwen-Image 1.0 workflows, so I'm wondering if we're going to see something similar with 2.1.
> The text rendering definitely is much, much better than anything else on the open weights market right now
Really? Because basically everything in those screenshots is completely garbled. I didn't follow it super closely but I thought Ideogram or whatever was really good for this particular use, with actual clear text.
This is my experience as well. Ideogram4 (assuming you are willing to put in the work to use the proper structured JSON input) is very accurate when it comes to text rendering in an image.
Regardless of the quality here, I see this as a failed business model. Subscribable libraries of skills / premade assets was effectively what Tailwind was doing, and they weren't able to make it work any longer.
I could see it maybe as a one-time purchase, but $90/yr/user is somthing I'd never grab.
Not to mention my $299 total Tailwind Plus lifetime license (was TailwindUI, then paid the difference for the Plus bundle) has like 100x more stuff in it than $90/yr or $149 lifetime for—check notes—11 UI transitions...
Don't get me wrong, I'm excited for a StarCraft shooter despite it not being StarCraft at all (I enjoy the universe). I wanted StarCraft: Ghost back in the day, and this feels like a taste of that dream.
That said, targeting a release of 2030 is wild to me. Look at the Astra game dev hype that's hitting twitter right now. Yes it's all filtered to the best possible examples / yes people are not one shotting these things, but still it's very clear that LLMs are beginning to be very capable at game dev in a way that doesn't look/feel like ass.
LLMs are going to become more capable over time. I think most of us can agree that eventually they'll be competitive with AAA developers given enough GPU cycles. I would argue that given rate of improvement, we're looking at that point coming in the next year or two.
It's just starting to be capable of making assets in blender / unreal. Given this rate of advancement, what is the state going to be like 4 years from now? It really feels a bit like the "travelling to distant stars" problem, where at some point it's faster to wait for better engines than to leave now. I worry that by the time this gets released, the market is going to be flooded with fully custom AAA games that are hyperniche. 4 years out is a long, long time to wait, and it's a longer time to bet on success / market dynamics.
I just read the book Play Nice: The Rise, Fall, and Future of Blizzard Entertainment by Jason Schreier (without having played any of their game or being aware that BlizzCon is around the corner). It seems they'll repeat their own mistakes.
I am getting more and more pessimistic about these studios and their AAA works. Every game is now a gamble with multi year effort with hundreds if not thousands of staff behind it. If a game doesn't sell, and that happens more than once, layoffs are coming. And for Blizzard specifically, it'd better be an online game that's going to be a multi billion dollar cash cow they can keep milking for years. If not, they would not work on the game in the first place.
Of course they have released many great games, but as the history has proven, you can't just keep coming up with new games that are original, fun if not addictive and also make money, especially with the fierce competition today.
Which makes me think I should spend more money/time on indie games than the action/adventure game from big studios that are increasingly repetitive and boring. We need more, smaller studios to give us something new.
Matches my experience. Indie games can be delightful, full of Easter eggs and heart that the dev(s) build for the fun of it. I just played through Hollow Knight for the first time and man, what a great experience.
I just started playing with building a game in Godot with Fable 5.1 last weekend and the pace is crazy. Obviously there's a world of difference between what looks playable on my machine and what a AAA studio with the best human artists and developers can put out, and I appreciate that they still exist and put in the first class effort, but... yeah. A 3-year timeline while this is going on seems out of touch with what's going on in the world.
Apparently 6 years is not an uncommon development cycle for AAA right now. For a random example, Ghost of Yotei which is also an open world game, was in development for 4 years. That game is a sequel too, and I think Sucker Punch is probably less dysfunctional than Blizzard at the moment. Still it might have been premature to announce it at this stage.
I'm not a gamedev so I might be wrong, but my friend showed me in 2020 a quick 30 minute tutorial in unity. In this time he made a basic 3d shooter game. The first video is the type of game a teenager would have made 20 years ago and the second one is of course a lot more impressive, but I believe the game engine is doing a lot of the heavy lifting there.
I had a bunch of extra Fable credits, so I spent about $10k running autoresearch loops on 200 of the top github repos with frontends to try and speed up the frontend performance. I distilled them down into a leaderboard of the most common wins, and built out an autoresearch loop that takes those learnings and applies them to your repo.
Works with your existing claude/cursor/codex sub in a cute custom TUI.
reply