Hacker Newsnew | past | comments | ask | show | jobs | submitlogin

> Wonder what am I doing wrong?

You're comparing 100b parameters open models running on a consumer laptop VS private models with at the very least 1t parameters running on racks of bleeding edge professional gpus

Local agentic coding is closer to "shit me the boiler plate for an android app" not "deep research questions", especially on your machine

 help



Looks at the headline: Qwen3.5 122B and 35B models offer Sonnet 4.5 performance on local computers

Yes and Devstral 2 24b q4 is supposed to be 90% as good but it can't even reliably write to a file on my machine.

There are the benchmarks, the promises, and what everybody can try at home


maybe a harness problem?

Having tried the Mistral Vibe harness that was supposedly designed for Devstral, that thing is abysmal. I feel sorry for whatever they did to that model, it didn't deserve it.

The thing I most noticed was asking it for help with configuring local MCP servers in Mistral Vibe - something it supports, it literally shows how many MCP servers are connected on the startup screen - it then begins scanning my local machine for servers running "MineCraft Protocol".

I want Mistral to do well, and I use their Voxtral Transcribe 2, that one has been useful. I'd even like a well made Mistral Vibe (c'mon, "oui oui baguette" is a hilarious replacement for "thinking"). But Mistral are so far behind, and they don't seem to even know or accept that they are.


The hardware difference explains runtime performance differences, not task performance.

Speculation is that the frontier models are all below 200B parameters but a 2x size difference wouldn’t fully explain task performance differences


> Speculation is that the frontier models are all below 200B parameters

Some versions of some the models are around that size, which you might hit for example with the ChatGPT auto-router.

But the frontier models are all over 1T parameters. Source: watch interview with people who have left one of the big three labs and now work at the Chinese labs and are talking about how to train 1T+ models.


> The hardware difference explains runtime performance differences, not task performance.

Yes it does.


Care to elaborate?

Certainly not Opus. That beast feels very heavy - the coherence of longer form prose is usually a good marker, and it is able to spit 4000 words coherent short stories from a single shot.

He's running a 35B parameter model. Frontier models are well over a trillion parameters at this point. Parameters = smarts. There are 1T+ open source models (e.g. GLM5), and they're actually getting to the point of being comparable with the closed source models; but you cannot remotely run them on any hardware available to us.

Core speed/count and memory bandwidth determines your performance. Memory size determines your model size which determines your smarts. Broadly speaking.


The architecture is also important: there's a trade-off for MoE. There used to be a rough rule of thumb that a 35bxa3b model would be equivalent in smarts to an 11b dense model, give or take, but that's not been accurate for a while.

> There are 1T+ open source models (e.g. GLM5),

GLM-5 is ~750B model.


Who would have thought ai labs with billions upon billions of r&d budget would have better models than a free alternative.

I'll add, AI Labs put a lot of resources into allowing the AI to search the web.. that makes a big difference

I use search as well via openwebui + searxng.



Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: