I have said this before as well: these top-of-the-line models write clever, convoluted code. The code looks intelligent from above, but is a maintenance headache. Makes entire thing fragile for future developments on top of it.
The smaller models, especially the aforementioned ones, they fail much more, but, do not write that insanity of the code. They do simple, non-clever coding like humans do. Much easier to maintain and build upon.
Qwen-3.6-27b is a wonderful model. Exceptionally good for it's size, and excellent in general as well. And with mtp available now, it can run at 60+ tps on a single 3090... this is roughly 30% faster tgs than most of the hosted ones being served from giant data-centers.
It keeps "branding" intact. This is similar to red-bull spending $3 BILLION annually on marketing... it's to keep the brand alive in the minds of folks and "associate" them as kinda the de-facto in given industry.
You can run the two agents in GAN like loop.. each trying to better the other. Give them a common task like design a better alternative to transformer model that uses max O(nlogn) memory, and the result comes closest to existing n^n implementations.
Good idea actually.. why haven't I tried this before.
As much as I absolutely love Mimo V2.5 Pro (it's a genuinely good model), I absolutely hate the way they calculate usage in their token plan.
For example: For a super small task in a small project that should not be consuming more than 500K total tokens after all tool calls included, their shown usage shot up to 152 million tokens.
But, when I scroll down on the same page, a table shows usage as 3 million tokens, out of which 2.5 million were cached.
This is such a huge conflict on the very same page. The bad thing is that the usage progress bar is shown against that 150 million token usage, not against that 3 million one.
This has been in discussions for at least past 3 months on reddit as well, and was precisely the reason I subscribed to their lowest tier, and for a single month only.
Update: their own harness, mimocode, shows total token usage as just 63.1K. We now have 3 entirely different values, differing in 3 orders of magnitude.
Update 2: So, I did the exact same task this time using DS4Pro, and total token usage was just 101K (as shown by opencode).
It was able to query csv, json, parquet and xlsx files, all locally. You could also mix files of different types in single session, and, manage multiple sets of files as individual sessions that you can switch easily whenever needed.
I built https://pagey.site initially for my own self since I was creating an insane lot of small one-page sites using LLM for various utilities/calculators/explanations etc. and needed a single place to host them.
Initial version was super minimal, but then, a lot of people started using it. So, revamped it into full product and launched it. Currently hosting over 300 sites, with 10 new sites getting added daily. All free for now. I personally has 53 sites as of now.
I loved the 'Build Software Like It Needs To Last' one. I feel like it's something that could be made into a poster and hanged where I could see it first thing in the morning.
This doesn't make sense. It completely misses that there are tons of other companies that will happily take the Nvidia capacity at the same price if their biggest customers reduce spend.
The smaller models, especially the aforementioned ones, they fail much more, but, do not write that insanity of the code. They do simple, non-clever coding like humans do. Much easier to maintain and build upon.
Qwen-3.6-27b is a wonderful model. Exceptionally good for it's size, and excellent in general as well. And with mtp available now, it can run at 60+ tps on a single 3090... this is roughly 30% faster tgs than most of the hosted ones being served from giant data-centers.
reply