The most prestigious law school admissions discussion board in the world.
|
 |
 
|
Review of IBM's new open weight model by a bigtech bro
Poast new message in this thread
Date: April 30th, 2026 10:54 PM Author: Jared Baumeister I had a shortish exploratory chat with Granite4.1:8b tonight.
It is a good model. The outputs feel something akin to deterministic which reflects the enterprise deployment IBM is shooting for. Input -> Output. The voice is pleasant and not overly beepboop robot. It hyperfixates on patterns (one message with a list kicks off 5 more) but it responses really well to faux system messages correcting it. The world knowledge is good and nuanced for an 8b model.
If I had a no-human-in-the-loop pipeline for evaluations or content parsing or something I’d 100% reach for Granite 4.1 first. It’s like,,, instruct tuned but only enough to accomplish its intended goal.
—— Oh! And! No reasoning! As an engineered design constraint. Neat! |
Date: April 30th, 2026 10:56 PM Author: i gave my cousin head who cares (http://www.autoadmit.com/thread.php?thread_id=5862289&forum_id=2#49856010) |
|
Date: April 30th, 2026 10:56 PM Author: oomox I understood like half of that. Explain it like I don't read hackernews (http://www.autoadmit.com/thread.php?thread_id=5862289&forum_id=2#49856011) |
 |
Date: April 30th, 2026 11:00 PM Author: Jared Baumeister It's designed to be fast and it was trained on high quality data. IBM really went their own direction (http://www.autoadmit.com/thread.php?thread_id=5862289&forum_id=2#49856026) |
|
Date: April 30th, 2026 11:02 PM Author: oomox What does "no reasoning" mean? (http://www.autoadmit.com/thread.php?thread_id=5862289&forum_id=2#49856029) |
|
Date: April 30th, 2026 11:09 PM Author: Jared Baumeister Everyone thought they needed reasoning models to get the best chat experience, but now with agentic harnesses the reasoning becomes a waste of tokens. I have it disabled in my Hermes Agent because it's too fuckin slow. IBM basically just saved us all from having to manually disable it (http://www.autoadmit.com/thread.php?thread_id=5862289&forum_id=2#49856064) |
|
Date: April 30th, 2026 11:10 PM Author: oomox What does that mean? Doesn't it need to do reasoning to follow instructions? (http://www.autoadmit.com/thread.php?thread_id=5862289&forum_id=2#49856068) |
|
Date: April 30th, 2026 11:28 PM Author: Jared Baumeister The model still has to follow intermediate steps during prompt processing, people are just finding more efficient ways to do it. One model (I forget which one) keeps the entire prompt intact at each layer. (http://www.autoadmit.com/thread.php?thread_id=5862289&forum_id=2#49856095) |
|
Date: April 30th, 2026 11:54 PM Author: oomox Yeah ofc, I just didn't know what "reasoning" meant and was too lazy to search (http://www.autoadmit.com/thread.php?thread_id=5862289&forum_id=2#49856136) |
|
Date: April 30th, 2026 11:33 PM Author: The Penis Like how chat gpt and claude for "hard problems" will go into a "chain of thought" where it "reasons". I turn it off for basic chats and only use it when I'm having it solve something difficult. Also sometimes it makes it stupider believe it or not. Also way faster without it. (http://www.autoadmit.com/thread.php?thread_id=5862289&forum_id=2#49856103) |
|
Date: April 30th, 2026 11:53 PM Author: oomox Got it, I thought it might mean that. I totally believe it could make it stupider in some cases. I'll turn that off on the ones I use and see if the results are better. (http://www.autoadmit.com/thread.php?thread_id=5862289&forum_id=2#49856134) |
|
Date: April 30th, 2026 11:31 PM Author: The Penis Would never use something that has "deterministic outputs" (http://www.autoadmit.com/thread.php?thread_id=5862289&forum_id=2#49856098) |
|
Date: April 30th, 2026 11:36 PM Author: $30 Kirkland brand winter coat I have news for you about women (http://www.autoadmit.com/thread.php?thread_id=5862289&forum_id=2#49856105) |
|
Date: April 30th, 2026 11:55 PM Author: JuanTwenty what tokens/sec are you getting (http://www.autoadmit.com/thread.php?thread_id=5862289&forum_id=2#49856142) |