Conversation

Notices

1. Embed this notice
   lainy (lain@lain.com)'s status on Friday, 24-Apr-2026 17:07:11 JST lainy
   there are now 3 open weight models that are reaching the quality of the closed ones (GLM-5.1, Kimi 2.6, Deepseek V4). There are two more that can write actually good code and small enough to run locally (Gemma 4, Qwen 3.6). This is not going away, no matter if OpenAI and Anthropic survive or not.
   • Embed this notice
     isutowaru (histoire3@shota.house)'s status on Friday, 24-Apr-2026 17:13:55 JST isutowaru

     in reply to

     @lain@lain.com Get useful toolGiant retarded companies dieBest ending

     lainy likes this.
   • Embed this notice
     cell classic (cell@shitposter.world)'s status on Friday, 24-Apr-2026 17:23:27 JST cell classic

     in reply to

     • isutowaru
     @histoire3 @lain free GPUs and RAM too to run local models
     lainy likes this.
   • Embed this notice
     lainy (lain@lain.com)'s status on Friday, 24-Apr-2026 18:13:03 JST lainy

     in reply to

     • h4890
     @h4890 any mac with > 32gb ram, those new unified memory systems by AMD work too. or a graphics card with at least 24gb vram (this is much faster).
   • Embed this notice
     h4890@merovingian.club's status on Friday, 24-Apr-2026 18:13:04 JST h4890

     in reply to

     @lain How much hardware dose it take to run those models at home with reasonable performance?

   • Embed this notice
     lainy (lain@lain.com)'s status on Friday, 24-Apr-2026 18:27:22 JST lainy

     in reply to

     • Harblinger
     @harblinger gemma 4 26b-a4b, qwen 3.6 27b and if it fits / you use a smaller quantization, qwen3.6-35b-a3b. they all run well in lm studio, if you don't already have a system for running them
   • Embed this notice
     Harblinger (harblinger@wizard.casa)'s status on Friday, 24-Apr-2026 18:27:24 JST Harblinger

     in reply to

     @lain have a favorite that can fit in 24gb yet? I feel like staying up a little late and trying one

   • Embed this notice
     lainy (lain@lain.com)'s status on Friday, 24-Apr-2026 19:02:34 JST lainy

     in reply to

     • rin
     @rin lm studio is the simplest i think. for coding things can be a mixed bag because that relies heavily on tool calls and for some reason nobody manages to support that properly on day one, so you often have to wait a week for the bugs to be ironed out
   • Embed this notice
     rin (rin@post.ebin.club)'s status on Friday, 24-Apr-2026 19:02:36 JST rin

     in reply to
     @lain how do you run the local models on macos? curious to try it out
   • Embed this notice
     Harblinger (harblinger@wizard.casa)'s status on Friday, 24-Apr-2026 19:37:49 JST Harblinger

     in reply to

     @lain cool thx :tipshat:
     I'll probably just do ollama + pi.dev https://docs.ollama.com/integrations/pi

     lainy likes this.
   • Embed this notice
     lainy (lain@lain.com)'s status on Friday, 24-Apr-2026 20:31:43 JST lainy

     in reply to

     • h4890
     @h4890 yeah, these small ones are really useable with normal hardware and they cna do real work.
   • Embed this notice
     h4890@merovingian.club's status on Friday, 24-Apr-2026 20:31:44 JST h4890

     in reply to

     @lain Interesting... that's nothing! I thought it would require at least several GPU cards and at least 20k USD or more in hardware investment.

   • Embed this notice
     Mogged To Death (demha@comp.lain.la)'s status on Saturday, 25-Apr-2026 00:24:23 JST Mogged To Death

     in reply to
     @lain ig geohotz side of the debate is winning lol
     lainy likes this.
   • Embed this notice
     Harblinger (harblinger@wizard.casa)'s status on Saturday, 25-Apr-2026 05:20:18 JST Harblinger

     in reply to

     @lain I'm impressed, it was able to write a (better?) kdenlive "earthquake" plugin than I got out of an old claude model when I first tested code generation

     Still couldn't get it to be keyframable, not sure why yet

     lainy likes this.
   • Embed this notice
     lainy (lain@lain.com)'s status on Saturday, 25-Apr-2026 05:21:14 JST lainy

     in reply to

     • Harblinger
     @harblinger the models are really getting better and more efficient, it's not just more compute and larger models.