Pekka Enberg
Verified account This account is verified. Learn more@penberg
Building databases for agents @Supabase. Previously Founder/CTO @tursodatabase (acq. Supabase), @ScyllaDB. Author of "Latency" (amzn.to/4nKI3Un).
1.2K Following 20K Followers 11K Posts
Latest posts 5
I have spent a lot of time and tokens trying to make Bonsai 2 27B work for coding tasks, but it seems people were right: ternary quantization is perhaps too lossy for this, since switching to full-precision Qwen3.8-27B improves performance quite a bit.
The frustrating part is that Bonsai often finds a bug fairly quickly and even identifies a possible fix, but then gets stuck in long thinking loops and can't decide what to do. I saw this both in ad hoc testing and on SWE-bench.
Oh well, the local coding adventure continues with a bigger model, I guess. Maybe I need to buy a machine with 64-128 GiB of memory also…
The frustrating part is that Bonsai often finds a bug fairly quickly and even identifies a possible fix, but then gets stuck in long thinking loops and can't decide what to do. I saw this both in ad hoc testing and on SWE-bench.
Oh well, the local coding adventure continues with a bigger model, I guess. Maybe I need to buy a machine with 64-128 GiB of memory also…
I started to use @superlogical Rex today and seems like a keeper!
What coding agents are people using with local models? I am seeing bunch of people reporting presumably good results with Qwen with 32 GB (or smalller!) machines. But when I tried it with my own harness, it would just enter an infinite thinking loop, which is apparently know issue…
Proud to be one of the many folks that was part in building the digital pyramid
Future programmers will look at the Linux kernel the way we look at the pyramids of Egypt, wondering, “How the hell did they build like that with the tools they had?”View quoted post on X
