> The results are ALWAYS terrible compared to CLaude Code.
How much effort did you spend to make the local setup work well?
Local models are not plug and play, you need to adjust settings to make them behave intelligently.
On my side one model looking super dumb right after install ended up being 'agentic' after using the right chat template, fixing how tool calls are setup, fine-tuning plenty of params values, and adjusting the whole setup to my own hardware. Claude.ai (free) was great at helping me sort that all out.
In my experience with Qwen3.6 35B-A3B, disabling thinking made the model generate inaccurate replies. Ask it for the recipe of egg salad and it gives you the recipe of an omelette.
Did I miss something, is it possible to have that model be reliable without thinking?
Turning off thinking mode you need a different tokenizer and a whole host of settings. It's more like switching your vehicle from gas to diesel, than turning your headlights from lo to hi.
I used the settings recommended on huggingface/unsloth's model page + whatever suggestion from various LLMs - didn't research too much myself which setting did what.
Frontier model labs don't know anything about good sampling settings, and why should they? Their belief is that "sampling doesn't matter, we can simply scale our model and improve logprobs quality that way"
Ignore whatever nonsense qwen says about what settings to use. Turn off top-p and top-k. Turn on top-n-sigma and keep it at its defaults.
See repetitions that you don't like? Ignore the "repetition, frequency, presence" penalties and turn on DRY (don't repeat yourself).
Want more creativity? Temperature is okay but XTC (eXclude Top Choices) is better.
Llama.cpp quietly has had far advanced samplers compared to the shit stuff Qwen et al recommends for years.
> Quantization-Aware Training (QAT) [...] allows preserving similar quality to bfloat16 while dramatically reducing the memory requirements to load the model
Pretty much. I plan on using this program for a long time. I don’t want a codebase that looks like something out of a H.P. Lovecraft novel when I have to fix something in the future.
Not everyone is paying for LLMs, even now. So I think it is perfectly reasonable to assume good intentions, here.
Someone spent their own tokens to ponder your code and thought they'd share the result. For anyone else looking, like me, I can see that this is probably going to come up relatively clean without having to spend my own tokens, or install it, and I'm more likely to, now that I can see that.
Which would have effectively overriden my whole bashrc config if I had blindly copy-pasted it.
A few minutes later, asking it to create a .gitignore file for the current project - right after generating a private key, it failed to include the private key file to the .gitignore.
I don't see yet how these tools can be labeled as 'major productivity boosters' if you loose basic security and privacy with them...
IME it’s a term that’s been popularized by generative AI solutions, a meme at this point, and doesn’t speak to real production readiness quantifiably (professionally). It’s something that I’ve seen models frequently claim during coding and planning sessions, and it can also be found around Reddit/Twitter/Github vibe coding spaces.
Seeing this term in marketing materials signals that the target audience is non-professionals (and I don’t mean this derisively, only that we need to apply a different lens).
> I'm not sure of what "production ready" is supposed to mean here
Given this text at the bottom:
> The high-performance HTML to PNG engine. Built for developers, agents, and automation. Completely free to use. All generated assets are public and ephemeral.
...I assume the implications are that:
1. this service will scale to meet request load without QoS degradation (i.e. it's probably running on FaaS infra), rather than being a fixed-size slowly-elastic cluster that would get choked out if your downstream service got popular and flooded it with thousands of concurrent requests
2. you can directly take the URLs the service spits out, and serve them to your downstream service's clients, without worrying much about deliverability, because there's an object store + edge CDN involved.
In other words, it's not just a single headless-chromium instance running on a box somewhere; you could actually use this thing as an upstream dependency and rely on it.
> the demo image is not optimized, `optipng` command decreases its size by 53.21%
Given that the author's imagined use-case is giving non-multimodal LLMs a way to emit visuals (the prompt at the bottom of the page starts "When asked to create visuals, charts, or mockups"), I think their idea is that the resulting rendered images would more-likely-than-not only be requested once, immediately, to display the result to the same user who caused the prompt to be evaluated.
Where, in that case, the metric of concern isn't "time+bytes cost for each marginal fetch of the resulting image from the CDN"; but rather "end-to-end wall-clock time required to load the HTML in the headless browser, bake the image, push it to the object store, and serve it once to the requesting user."
OptiPNG would slightly lower that last "serve it once" cost, but massively inflate the "bake the image" time, making it not worth it.
(I suppose they could add image optimization as something you could turn on — but "image optimization at the edge" is already a commodity product you can get from numerous vendors, e.g. Cloudflare.)
It might be worth considering a feature to time/schedule each flow's animation, rather than having them run in an infinite loop, all at the same time.
UX feedback:
* The animation and the whole interface are sluggish on firefox/linux. There's about 1 sec delay after each action (like clicking on an option).
* The site's CSS does not load on an old version of Chrome - v90 - (and the chart and animation don't either).
How much effort did you spend to make the local setup work well?
Local models are not plug and play, you need to adjust settings to make them behave intelligently.
On my side one model looking super dumb right after install ended up being 'agentic' after using the right chat template, fixing how tool calls are setup, fine-tuning plenty of params values, and adjusting the whole setup to my own hardware. Claude.ai (free) was great at helping me sort that all out.
reply