nero

[nero @jmnilabs]

@0xSero What engine? Those are great numbers and I have yet to see someone post about TP2 DS41 1017 yet. I have 4 256GB ordered, they just take so long to arrive.

[nero @jmnilabs]

@alexocheema @exolabs 220 C1? What engine?

[nero @jmnilabs]

@joelniklaus @deepseek_ai @huggingface Only fools would believe that Hermes is a better coding agent than Codex or Pi. Whatever "task" this is based on is not accurate.

[nero @jmnilabs]

@aijoey @QwenDevs @Alibaba_Qwen @QwenAIFans @AlibabaGroup Which Qwen4 is this, if you know?

[nero @jmnilabs]

@hxiao Not true. Look at the updated numbers.

[nero @jmnilabs]

@niklaslenz_ai Hi. Enjoy significantly better speeds with this! Uses LIL/NVIDIA's QAD checkpoint and their hyperflare engine. As is that will work with two DGX Sparks! Code at 85 t/s and prose 62! No loss in quality for the speed. https://t.co/CWjwCJ4pN6

[nero @jmnilabs]

@paulmarin90 @0xSero Do yourself a huge favor and run https://t.co/U9gMiTfgm4 from LIL. You will have significantly better performance! The work is top notch and serves the community, not for views and is a full up nonprofit!

[nero @jmnilabs]

@christopherowen Stolen work from the folks at LIL.

[nero @jmnilabs]

@CerebralCoding_ Seems like a step up in capability, nice. How does it fair at larger contexts like 128k? What bench method did you use to get your decode numbers?

[nero @jmnilabs]

@TechMDAI I'll take the more efficient Macs. RAM is getting pretty valuable and models are only getting bigger. Also software is getting better as we speak.

[nero @jmnilabs]

@someRandomDev5 @superalesha Its for the views. Its been out since day 1 haha.

[nero @jmnilabs]

@ivanfioravanti I'd love to see this on two M5U!

[nero @jmnilabs]

@Orwelian84 @ivanfioravanti The developers are @YourLocalAILab !

[nero @jmnilabs]

@MiaAI_lab https://t.co/Z3Z483lX6o

Is already a thing! Works great.

[nero @jmnilabs]

@sfxnz @MiaAI_lab I do this with a fold8. It was a breath if fresh air moving from the apple locked ecosystem to something significantly more manageable.

Using the fold8 for agentic work is really great.

[nero @jmnilabs]

@TechMDAI Its engrams quantized. You'll see no changes other than MAYBE from offloading to ram vice ssd.

[nero @jmnilabs]

@ivanfioravanti @chrisfzz2 Need someone to test it!

[nero @jmnilabs]

@MiaAI_lab @exploit_intel This is an early works engine. Performance already improving!

[nero @jmnilabs]

@T_Wrex_Baby It would be great for M5U if you’re seeing those numbers you’re claiming. I suggest you join the discussion more academically and post some of the results you’re achieving. I for one have an M5U on the way and would love to see it perform as well as you say.

[nero @jmnilabs]

@mxfp0 These numbers are pretty slow. 4 sparks gets you in the ballpark of 60-65 c1, and 100-120 c1 code. Prefill 4k. Check out the work and use @YourLocalAILab’s vLLM and b12x! Much more optimized.