Paul

COO, coach and the guy that gets things done. Likes are simply a way to store a reference to content/posts, and do not represent me liking any content.

[Paul @pbastowski]

@ivanfioravanti I have tried every dev and rc release and still going back to 0.3.12 for serious work - the fastest and most stable.

[Paul @pbastowski]

@antirez @liuliu I found this to happen sometimes. It seems that it’s some sort of context pollution that makes the model suddenly retarded.

The only solution I found was to start a brand new session and continue with the same exact checklist item that failed, and it would then work perfectly.

[Paul @pbastowski]

@jun_song @grok oMLX LLM runner’s prefix caching makes agentic/batch runs almost instant between steps. No advantage to the spark at all.

[Paul @pbastowski]

@hirenpatelatl @jun_song @grok M5 Max also wins in batch if you use oMLX as the LLM runner for agentic coding, due to prefix caching. Batch is then almost instant.

[Paul @pbastowski]

@Prince_Canuma @Alibaba_Qwen Great news!
Does thinking mode have to be on for the soles gains?
Also, what TPs have you seen with thinking mode off?

[Paul @pbastowski]

@ivanfioravanti @Kev96790724 @Teknium I have a separate MEMORY.md file, which keeps the "game save" of the current situation and gets cleaned up once a while, with the stale entries being moved to ARCHIVE-yyyy-mm-dd.md files. I found this to work best for me.

[Paul @pbastowski]

@camilobayarri @jun_song The can be ok, if you don't mind slow inference speeds.

[Paul @pbastowski]

@camilobayarri @jun_song Assuming it’s for agentic coding and vibe coding then Qwen3.6 27b or qwen3.6 35b a3b on an m4 max or m5 max MacBook with at least 64gb ram. 35b model will fly at q8 even and 27b model will be more clever, but much slower.

[Paul @pbastowski]

@Youssofal_ Nice!

[Paul @pbastowski]

@RaminNasibov Blue Max on C64

[Paul @pbastowski]

@jun_song Haha, yeah, same here. I had to reduce the hours drastically, though. The body can only take so much focused learning.

[Paul @pbastowski]

@dom_lucre Good voice. He looks older than 15 though, doesn’t he?

[Paul @pbastowski]

@fentflipEUW @silentlink1 Where did you find the 40gb EU data option in the flex app?

[Paul @pbastowski]

@bridgemindai Have you tried qwen3.6 35b a3b q8 mlx running on oMLX?

You'll get 90 tps at low context and 64tps at 70k context, and it will feel faster than gpt at times.

Granted, it's not as clever, but it is a very good and fast coding model if you instruct it well what to do.

[Paul @pbastowski]

I saw that months ago while building a mobile app with Codex and Gpt5.3. Gpt was surgical, precise and only did what I asked. Opus decided it knew better than what I asked, and implemented what it wanted instead. I would ask a question and it would start coding. That was really annoying.

Gpt > Opus

[Paul @pbastowski]

@hicasamadim 72

2x2 x 2 =8
5x5 x 2 =50
6x6 x 2 =72

[Paul @pbastowski]

@antirez @badlogicgames @FrameworkPuter Half power and tps is indeed not perfect. I’ll try —power 75 in a moment and see how my fans like it.

[Paul @pbastowski]

@antirez @badlogicgames @FrameworkPuter Selecting low power also turns off the fan for half the tps.