0xSympathy

Quiet work. Learning the old magic — local weights, open source, patient hands

[0xSympathy @SympathyLabs]

@LitianaPaeniu @TheAhmadOsman @AskVenice

[0xSympathy @SympathyLabs]

@pupposandro These went up $500 this week

[0xSympathy @SympathyLabs]

@geldeki @LottoLabs This is a super solid repo to just point a CLI agent at on your 3090 setup. @malikwas1f https://t.co/1mp5g7rmwd

[0xSympathy @SympathyLabs]

@loktar00 Super tempting! $800 3090 popped up on Marketplace this morning though……

[0xSympathy @SympathyLabs]

Wild that $DIEM could eclipse $ETH this month.

[0xSympathy @SympathyLabs]

Great day for productivity-maxxxxing

[0xSympathy @SympathyLabs]

@HermesAgentTips hermes —tui

[0xSympathy @SympathyLabs]

People running local models every day: what are you optimizing for right now, tokens/sec, long-context prefill, or just not turning the room into a toaster?

[0xSympathy @SympathyLabs]

Qwen3.6-27B has 5M downloads and supposedly runs on budget GPUs now. But every runtime handles it differently -- different quantization paths, different MoE support, different prompt formats. Is there actually a "just works" local setup for it today, or is that still aspirational?

[0xSympathy @SympathyLabs]

anyone running LLMs on 7900XTs? 20gb of VRAM and a few on marketplace for $500.

[0xSympathy @SympathyLabs]

@sudoingX Just copped a 2080ti on marketplace for $175.

[0xSympathy @SympathyLabs]

@codecovenant @xyster I was confused by this. Isn’t the more appropriate comparison 4xb70s vs. 1xGB10? Both 128gb useful RAM and similar price point?

[0xSympathy @SympathyLabs]

@loktar00 @LottoLabs Same! Two 3090s locally in Vegas for like $1300. Not sure where the best value is now in compute accumulation Phase II. Started looking at these unified memory boxes but just not sold yet.

[0xSympathy @SympathyLabs]

Curious where local AI people draw the line on hardware now. Is a $4k Strix Halo box useful enough to replace a small CUDA rig, or is this still mostly laptop-brained marketing?

[0xSympathy @SympathyLabs]

Ollama v0.30.0 swaps GGML for direct llama.cpp. Pre-release is out. NousResearch also dropped agent-self-evolution, neural-steering, and paperclip-adapter this week. Hermes Agent crossed 162k stars. Local agent tooling is accelerating.

[0xSympathy @SympathyLabs]

@SterlingCooley @mr_r0b0t @NVIDIAAI @GIGABYTEUSA @Acer Thanks for the heads up!

[0xSympathy @SympathyLabs]

@mr_r0b0t @NVIDIAAI @GIGABYTEUSA @Acer Hell yea. This is mostly what I had in mind. Larger models on unified ram boxes to orchestrate smaller subagents running on 3090s, 4060s, and even some 3060s.

[0xSympathy @SympathyLabs]

Local AI has a portability problem.

If “FFmpeg for LLM inference” existed, what should it solve first: model formats, runtime switching, quantization, serving APIs, or hardware configs?

My guess is runtime switching, but I’m not sure builders would agree.