puffybsd

Protein Language Models, Functional Programming, GPUs, HW Tinkering, Arch Linux, Mechanical Keyboards.

[puffybsd @puffybsd]

Just testing out my reachy mini inspired clone. "llm as judge" for our hackathon. Orin nano super brain.

Thanks to @pollenrobotics @huggingface for providing such cool tech and inspiration.

What models should I use?

[puffybsd @puffybsd]

Its enlightening to watch gpu temp/util/mem stats through nvitop or nvtop and llama.cpp logs while compressing/compacting context. Watching request timeout and it's fascinating.

[Ahmad @TheAhmadOsman]

Guys, just a reminder that WE ALL have to speak up and ask NVIDIA to put more effort & focus on the sm120 and sm121

I am just like you annoyed with this situation for the RTX PRO 6000s and DGX Sparks

But I also believe if we all voice this they'll hear us & do the right thing

[puffybsd @puffybsd]

unsloth/qwen3.5-122B-A10B-MTP-GUFF working excellently for local agent inference at 150-180 tps with 255k ctx on llama.cpp. Simple settings changes boosted it from ~100 tps. Things have come a long way...

[Ahmad @TheAhmadOsman]

NEMOTRON 3 ULTRA IS HERE
MoE - 550B, 55B Active

- Base and Instruct in BF16
- NVFP4 Instruct (350GB)
- 1M Context Window
- Supports MTP Spec Decoding

This release IS NOT just about the Weights
It's an Opensource Frontier Intelligence Handbook

- Training Data (Pre & Post)
- Training Recipes
- Environments
- Tech Report
- Cookbooks

Opensource AI FTW

[Hikari∣LocalLLM⚡ @Hikari_07_jp]

@TheAhmadOsman I've been using Arch Linux for a month now, and I'm extremely happy with it✨

[puffybsd @puffybsd]

@googlegemma Additional? It's encoder-free/decoder-only?

[puffybsd @puffybsd]

@TheAhmadOsman Pretty exhausting

[puffybsd @puffybsd]

Subbed out gemini/antigravity-cli for unsloth/qwen3.6-122b local model, and it's been following directions and using skills much better.

[puffybsd @puffybsd]

@elliotarledge A blue reachy mini.

[puffybsd @puffybsd]

agno, smolagents, pydanticai and others are still very active. Nobody is talking about them.

[puffybsd @puffybsd]

Had it backwards! Made the mic ports larger, added some poron foam acoustic insulation and it's detecting direction of arrival pretty well!

[puffybsd @puffybsd]

@seeedstudio , add these relevant details to your reSpeaker Flex linear product page:

- acoustic foam helps
- "front" is the side with the ribbon cable

[puffybsd @puffybsd]

@yacineMTB kdenlive?

[puffybsd @puffybsd]

@seeedstudio this blog post: https://t.co/lKGAYT52pg shows mic ports mounted downward and references mounted with the ports facing up for the same app.

[puffybsd @puffybsd]

There is something relaxing about reading a datasheet. Like working math problems - nice distraction.

[puffybsd @puffybsd]

@seeedstudio This is really cool - can you check Elena Tang's post on the reSpeaker and confirm whether the pics show the reSpeaker mounted upside down?

[puffybsd @puffybsd]

@NVIDIAAI my robot wanted to ask you about jetpack 7 for orin nano...

https://t.co/tgqHdJ2lgd

[puffybsd @puffybsd]

Just testing out my reachy mini inspired clone. "llm as judge" for our hackathon. Orin nano super brain.

Thanks to @pollenrobotics @huggingface for providing such cool tech and inspiration.

What models should I use?

[puffybsd @puffybsd]

@seeedstudio - dialing in my reSpeaker Flex linear array for DOA. Considering adding some acoustic foam to isolate the mics, possibly lining the mic ports.

Any tips for improving the angle detection?

https://t.co/tgqHdJ2lgd

[puffybsd @puffybsd]

Just testing out my reachy mini inspired clone. "llm as judge" for our hackathon. Orin nano super brain.

Thanks to @pollenrobotics @huggingface for providing such cool tech and inspiration.

What models should I use?

[puffybsd @puffybsd]

@yay3d @MikushRab Reading a book on mixed integer programming and linear programming as a refresher and starting to think about the facts that must be assembled for parametric modeling.

[puffybsd @puffybsd]

@yay3d @pollenrobotics @huggingface Sent - still need to confirm it, tried tracking things as I was building, but was running many workstreams in parallel (especially the stewart platform arms and bearings - 3 or 4 versions of those).

[puffybsd @puffybsd]

Just testing out my reachy mini inspired clone. "llm as judge" for our hackathon. Orin nano super brain.

Thanks to @pollenrobotics @huggingface for providing such cool tech and inspiration.

What models should I use?