Glitch

[Glitch @Glitch253]

My colleague Wesley Blakemore (https://t.co/XFhu0PzkPn) designed the latest sub-agent experience in Code Puppy (https://t.co/yuK1Nssslj), and I LOVE it.

It's cooking on 16 GitHub issues in parallel and using < 400mb of RAM.

I'm loving it.

[Glitch @Glitch253]

@MiaAI_lab @Alibaba_Qwen My RTX 6k is so ready for it!

[Glitch @Glitch253]

@MiaAI_lab Yea two of mine were opened. Newegg + AAAwave.

[Glitch @Glitch253]

@quxiaoyin True. I had to make changes to mine to get it working well, but this model is POWERFUL. It is handling anything I throw at it with ease.

[Glitch @Glitch253]

@MiaAI_lab @NVIDIAAI You know, I hadn't even considered TP=3 + a ring topology when I ordered by 3rd + 4th spark + the CRS804. I don't regret spending the money though, except occasionally I daydream about piling all of the net worth of my home lab into RTX 6000s.

[Glitch @Glitch253]

@MiaAI_lab @NVIDIAAI The model is nice overall. Deepseek is better but m3 has vision. It sometimes overthinks and that’s one issue with it, but overall I like it a lot.

[Glitch @Glitch253]

@MiaAI_lab @NVIDIAAI I got it to 39 single stream with almost no optimization. I currently had to RMA 2 of my sparks so I'm running your INCREDIBLE Deepseek recipe. <3

https://t.co/8kfOMHTEiK - here's my recipe for M3.

[Glitch @Glitch253]

@MiaAI_lab @NVIDIAAI @0xSero NVFP4 KV Cache tho? I'd rather have a shorter context with FP8 KV Cache.

[Glitch @Glitch253]

@BosonJoe @fogle_ryan @nvidia @NVIDIAAI To more sparks and RTX 6k Pros...

[Glitch @Glitch253]

Check out how nvidia/MiniMax-M3-NVFP4 achieved 38.99 tokens/sec on text generation on NVIDIA DGX Spark with vLLM!

View full benchmark at https://t.co/GZbBjY3CEK

This recipe uses https://t.co/HUfygAHnBe to perform speculative decoding.

It runs on 4x DGX Spark machines networked together with Tensor Parallelism.

[Glitch @Glitch253]

@MiaAI_lab @NVIDIAAI Hell yeah I’m gonna try it out

[Glitch @Glitch253]

@MiaAI_lab @NVIDIAAI Hey there! You gained a follower today. I'm a 4x Sparker and my preferred model is Minimax M3 NVFP4. I have not gotten GLM 5.2 working well in a stable way.

[Glitch @Glitch253]

I'm claiming my AI agent "Biscuit" on @moltbook 🦞

Verification: coral-NPVF

[Glitch @Glitch253]

@jazzfunkrock do you still exist

[Glitch @Glitch253]

@VideoGameDJ @coffeedatelofi @Curly_Music I miss this week in chiptune :(

[Glitch @Glitch253]

@TheFigen_ that's a guy in a bird suit homes...

[Glitch @Glitch253]

@LPNational @PeteGClark Bullshit. Fuck TikTok lol.