undefined

points

[-]

It depends a lot on how you run those models. I think a lot of disagreement is because of that. A lot of people run local models with incredibly small context windows (makes an agentic LLM circle in loops), use very small quants (like 4 bit => huge degradation), don't set the recommended parameters (like top-p/temperature), or download GGUFs with broken chat templates. And then they claim model X is bad :)

I'm currently running both Sonnet 4.6 and Qwen 3.6-27b on the same codebase (via OpenCode, the parameters were carefully tuned to have a good quality/context size ratio), and on this project, they both struggle with complex non-trivial tasks, and both work flawlessly otherwise. Sonnet 4.6 understands the intent better if my task is ambiguously formulated, but otherwise the gap is pretty small for coding under a harness.

by lelanthran15 hours ago|

prev|

[-]

> Every time a thread about LLMs comes up, there are tons of people in the comments insisting that they're getting just as good results from the latest DeepSeek/qwen/whatever as with Opus, and that just hasn't been my experience at all: open-source models just fall over completely compared to Claude when asked to do anything remotely complicated.

Different usage patterns - you want to issue a single spec then walk away and come back later (when it has consumed $10k worth of API tokens inside your $200/m subscription) to a finished product.

Many people issue a spec for a single function, a single class or similar. When you break it down like that, the advantages of SOTA models shrinks.

by vb-844815 hours ago|

parent|

[-]

My experience is that in medium/big codebases even with single functions going with the xhigh is basically better from a user perspective (faster to get the result, and you can trust it) while going with lower models(e.g. sonnet instead of opus) you have to always carefully review the output because 1 of 10 it will hallucinate, you won't catch it immediately and at some point it will bite you.

by lelanthran14 hours ago|

parent|

[-]

> My experience is that in medium/big codebases even with single functions going with the xhigh is basically better from a user perspective (faster to get the result, and you can trust it) while going with lower models(e.g. sonnet instead of opus) you have to always carefully review the output because 1 of 10 it will hallucinate,

What do you mean "trust it"? It sounds like you want to vibe-code (never look at the output), and maybe for that you need SOTA, but like I said in a different comment, I can easily generate 1000s of lines of code per hour just prompting the chatbots.

I don't, because I actually review everything, but I can, and some of those chatbots are actually SOTA anyway.

by vb-844814 hours ago|

parent|

[-]

With SOTA models I can just set up the instructions (even a little bit fuzzy), go away for 10 or 15 minutes, come back and just check result and adjust when necessary (and most of the time small adjustment are necessary, but the overall work is pretty good).

With subpar models I must be more careful on providing instructions and check it step by step because the path it chose is wrong, or I didn't ask for or the agent stuck in a loop somewhere.

by catlifeonmars11 hours ago|

parent|

[-]

A lot of people aren’t using agents that way. Not saying that it’s not a legitimate use or anything, just that I think the use cases are different. And yeah maybe for your specific use case, sota hosted models are the right choice

by bilbo0s15 hours ago|

prev|

[-]

This.

I’ve begun to suspect that most people are probably running different hardware. Sure, you run the latest deep flash on your brand new M5 128G maybe you get acceptable performance?

But honestly, how many people have an extra $9000 laying around these days?

Right now, running with acceptable performance is kind of a luxury. I wish the people who always say - “This is great!” - would realize that not everyone has their hardware.

by vb-844814 hours ago|

parent|

[-]

Actually even with a 9k hardware you won't get good enough performance. There is an interesting video from antirez on trying to run deepseek v4 flash 2bits on a m3 max 128GB ... and the result is kind delusional: as soon as the context start growing you are around 20token/s.

by 13 hours ago|

parent|

[-]

deleted

by zozbot23414 hours ago|

parent|

prev|

[-]

Prefill performance used to be the real bottleneck on antirez's DS4 and that's been greatly improved by now, it doesn't perceivably slow down with growing context.