A software developer working on https://xitter.cf/t.co/DzOih8KEW2 and https://xitter.cf/t.co/e4f8vy1OlT https://xitter.cf/t.co/kWDRC13CXw
Joined May 2008
- Tweets12K
- Following422
- Followers2.1K
- Likes8.1K
I go away for a week, come back and now due to open source work can run GLM 5.3 Flash, a 320B param model on my desktop at 35 t/s with 2200 t/s prefill.
Just buy some hardware and wait seems to be a decent proposition these days.
boyter retweeted
Breadcrumb is a computer flight recorder and context manager for your AI. innerloop.works/breadcrumb
Great. Watch the government speed run us to a backwater over this. Buy hardware and run your own models. abc.net.au/news/2026-09-24/a…
Noticed a lot of agents were hitting rate limits on searchcode.com so I have raised them... considerably. Watching the server now, but all seems OK... so far.
boyter.org/posts/did-i-just-… Nice performance improvement to scc with me uncovering some issues.
The next version of scc, is cooking when it comes to performance. v4.0.0 was all about intelligence, v4.1.0 is all about the speed. Also accuracy fixes, but mostly speed.
Every day Warhammer 40k becomes more realistic. We are what 60 years away from nobody knowing how code works.
According to this badge github.com/boyter/scc has hit over 1 million downloads. Not going to lie, I thought it would have been more!
Its kinda wild to me that you can run close to frontier models at home on... expensive, but not stupidly priced hardware.
If you're into Local AI, running models on consumer GPUs, you really need to be paying attention to what's happening right now.
The Artificial Analysis Intelligence Index has been updated to include Qwen3.8-Flash-Next.
Qwen3.8-Flash-Next - 56
Qwen3.8-27B (xhigh) - 53
Both of which i'm running on my 3090.
Smaller Q4/Q5 quants, running entirely on my own hardware. The results i'm getting from these models are impressive, insane.
Way beyond what i was able to get locally before these new model releases.
If you have a 3090, then you need to check my other posts, you're seriously missing out on lot of goodies.
Start incorporating local AI for some tasks and see for yourself.
Dead internet theory really is a thing. The stats for my blog for the last few days. Noticed a surge of traffic this morning, which as far as I can tell is down to bots mostly.
I did a thing. Added a Arc B65 to go with a 5090. B65 runs Qwen3.8-27b at 42 tokens per second. Runs full time allowing me to test things on the 5090 and dual wield when I need more power.
Atlassian launches a agent harness (because everyone needs one) teamworkgraph.com/cli but apparently vibed the site because the top bar does not work.
@Atlassian might want to fix that, you are making all Aussies look bad.
Latest release of scc v4.0.0 with a blog post, as it starts moving into the code intelligence space boyter.org/posts/sloc-cloc-c…
Since I have a nVidia 5090 and B65 Arc card... im testing splitting models across both. Which sounds insane, but works. GPT-120B running split on both getting about 40 tps.
More things to try here, but the current state of large models is more affordable than you think.
Well im rich, goodbye losers whom I've always hated!
Seriously though if I got this level of support from every download... id actually be able to work on it full time for the rest of my life at this point.
Was tracking some referrals from fossil scm to scc, found out its linked on it fossil-scm.org/home/doc/trun… in its git comparison. Cool!