Capicua

[Capicua @Capicua25x]

@theo My customer's data does not leave my local servers and that has a value.

[Capicua @Capicua25x]

@bertholomusai He should switch to open models. Why leave your company exposed to such risks by being beholden by another? AI sovereignty is the way.

[Capicua @Capicua25x]

@MiaAI_lab OpenCode v2 is great. I like it better than DSH.

[Capicua @Capicua25x]

@heinenbros This is a wake up call for you to invest in hardware and run open models. The open models of today are as good as Opus 4.8, and your business was running that just a few months ago just fine. DeepSeek 4.1 Flash, GLM 5.3 Flash, and Qwen 3.8 Flash Next are simply amazing.

[Capicua @Capicua25x]

@jun_song It would be reckless for me to buy 2 more Sparks to get TP4 at this point. I will survive with DeepSeek 4.1@3bit Flash on my Dual Sparks and Qwen 3.8@4bit Flash on my dual R9700 + 128gb of system ram. If newer models can't run later on, then so be it.

[Capicua @Capicua25x]

@jdpcasey @0x0SojalSec Maybe for 27B, but Qwen Flash is the real deal.

[Capicua @Capicua25x]

@bertholomusai @MiaAI_lab I've been doing hours long, and it now self recovers like the API does. No issues yet.

[Capicua @Capicua25x]

@bertholomusai @MiaAI_lab I'm running into reasoning loops and filed issues on your git with reproducers. I've confirmed that it's still happening on 0.5.

[Capicua @Capicua25x]

@0xSero Amazing

[Capicua @Capicua25x]

@jun_song I didn't have the money to buy two more, so I'll just have to make do until 2028 or whenever prices start to come down.

[Capicua @Capicua25x]

@jun_song OpenCode v2 has fixed all nearly the issue I had, and unlike DSH, it has a native CLI as well. The mobile web UI is now good.

[Capicua @Capicua25x]

@bertholomusai @deepseek_ai The current state of GLM 5.3 Flash is that the fast lane is tied to a NC license, and concurrency is locked to 1 on Tensorfold for MTP, so this does what I need very well.

[Capicua @Capicua25x]

@AWar1586398 @bertholomusai @ashxhart DeepSeek 4.1 is 4bit native. There's no fp8.

[Capicua @Capicua25x]

@bertholomusai Thanks for the awesome work. I'm rebasing now.

[Capicua @Capicua25x]

@Tech2Wild I have 2 and was planning on buying 2 more in February, but now I'm locked out.

[Capicua @Capicua25x]

@MiaAI_lab DeepSeek 4.1 on 2x Sparks. If you can unlock the frame buffer ram, maybe we can get concurrency and 1M context.

[Capicua @Capicua25x]

@MiaAI_lab @AnthropicAI On the other side of the coin of the knowledge used to secure a site is also the knowledge to attack. They're worried that you can trick Opus into doing the later. We're really going to have to use open models to help us secure our sites.

[Capicua @Capicua25x]

@MiaAI_lab DSpark for 5.3 by RedHatAl is out. Which means we now have a speculator that is for commercial use: https://t.co/E3cwSUPfJD

[Capicua @Capicua25x]

@MiaAI_lab @BrandonMusicKy Upstream licenses *can* change to something unfavorable. This is one way to freeze it, and it's allowed.