opencode on Nemotron-3-120B:
-
@aegrumet @theDanielJLewis Claude 4.6 is my daily driver. Just trying out opencode with a local model for comparison. A-B asking the Claude and Nemotron-3-120B to build the same thing side by side to compare.
It feels like there is a gap between large models and truly giant models. The difference between a good 9B model and a 120B param model is very noticable. But to get to the next noticable jump in accuracy/tooling you have to go to a 1T or larger model.
@theDanielJLewis @aegrumet I could be wrong. It's just my observation based on what I've seen so far.
-
@aegrumet @theDanielJLewis Claude 4.6 is my daily driver. Just trying out opencode with a local model for comparison. A-B asking the Claude and Nemotron-3-120B to build the same thing side by side to compare.
It feels like there is a gap between large models and truly giant models. The difference between a good 9B model and a 120B param model is very noticable. But to get to the next noticable jump in accuracy/tooling you have to go to a 1T or larger model.
@dave @aegrumet @theDanielJLewis
As a sloperator, got 5.3 and Claude 4.6 have felt pretty equal. The difference in token cost however has been a factor 20x higher on Claude. Costs are calculated on using opencode's payment system.
-
5.4 codex not available yet, you mean 5.4 general better for coding than 5.3 codex ?
-
@dave @aegrumet @theDanielJLewis
As a sloperator, got 5.3 and Claude 4.6 have felt pretty equal. The difference in token cost however has been a factor 20x higher on Claude. Costs are calculated on using opencode's payment system.
AI tools are evolving faster than human apes, and you can cross-check one AI tool with another AI tool, so slop-slur is not justified in using AI tools for software development
fun exercise: ask your AI chatbots: “is there a God? Answer yes/no inky”
BTW: when promised interview to me aboot vibe coding?
-
5.4 codex not available yet, you mean 5.4 general better for coding than 5.3 codex ?
@csb Yes, that seems to be how it's playing out.
-
@dave @aegrumet @theDanielJLewis
As a sloperator, got 5.3 and Claude 4.6 have felt pretty equal. The difference in token cost however has been a factor 20x higher on Claude. Costs are calculated on using opencode's payment system.
-
@dave @aegrumet @theDanielJLewis
As a sloperator, got 5.3 and Claude 4.6 have felt pretty equal. The difference in token cost however has been a factor 20x higher on Claude. Costs are calculated on using opencode's payment system.
@adam @dave @aegrumet @theDanielJLewis I would suspect the reason you're seeing higher toking usage in Claude is because the the maximum context window for gpt 5.3 is 400k where the maximum contacts window of Claude 4.6 is 1 million. Therefore, it has the ability to send a whole lot more information per prompt request. The beast is hungry and it loves to be fed!
-
@aegrumet @theDanielJLewis Claude 4.6 is my daily driver. Just trying out opencode with a local model for comparison. A-B asking the Claude and Nemotron-3-120B to build the same thing side by side to compare.
It feels like there is a gap between large models and truly giant models. The difference between a good 9B model and a 120B param model is very noticable. But to get to the next noticable jump in accuracy/tooling you have to go to a 1T or larger model.
can you check next time you have free time how long does it take your Nemotron-3-120B LLM to answer to this prompt (I want to compare to my Nemotron):
I have three shirts. I need 3 hours to dry them outside in the sun. How long does it take to dry 30 shirts?
-
@dave thanks, very similar time
-
by the way: since @adam doesn't want to give me interview for my YouTube channel about vibe coding despite promising it, if you're interested I can interview you Dave - I'd send you zoom link and you would share your screen and just talk and demonstrate, just like I did with Big Loaf from North Carolina here https://www.youtube.com/watch?v=zv_XAyDU0qo - you would hear my real voice which I would change with AI voice changer later on before publishing on YouTube. Let's not overthink.
Hello! It looks like you're interested in this conversation, but you don't have an account yet.
Getting fed up of having to scroll through the same posts each visit? When you register for an account, you'll always come back to exactly where you were before, and choose to be notified of new replies (either via email, or push notification). You'll also be able to save bookmarks and upvote posts to show your appreciation to other community members.
With your input, this post could be even better 💗
Register LoginWelcome To Podcasting.Chat!
This forum is for podcasters, podcast guests, and podcast enthusiasts alike to share tips, tricks, and their love of the medium.
This forum is fully federated, so you are able to contribute to any discussion here through your own software of choice (e.g. Mastodon, Misskey, Lemmy, Piefed, etc.). So you can sign up for an account here and it federates around the Fediverse. You can also follow feeds and topics from your other Fedi-enabled accounts.
Recent Posts
-
VICE Studios builds out leadership team #pressrelease #podcasting https://podnews.net/press-release/vice-studios-execs
-
@alecksgates I’m seeing that Iroh uses a debug/tracing crate. I’ve enabled RUST_LOG for that and we’ll see if that catches whatever is stalling the channel.
-
Nomono Unveils Audio/Video Sync at SXSW to Expand Into a Full-Stack Creator Workflow #pressrelease #podcasting https://podnews.net/press-release/nomono-audio-video-sync
-
Nomono Appoints Owen Grover as General Manager, US #pressrelease #podcasting https://podnews.net/press-release/nomono-owen-grover
