AI enabled mindset with passion for health and wellness especially in kids.
Earth.Planet
- Tweets 421
- Following 606
- Followers 109
- Likes 835
Thomas Eastham retweeted
Concurrency rerun with TensorFold 0.6.1 on the same Spark.
16 at once: 62 → 372 tok/s
First token at 16 at once: 117s → 0.3s
128K prompt read: 882 → 1,515 tok/s
Same NVFP4 weights and DFlash2 drafter, temp 0, thinking off. 0.6.1 ran with --parallel 16
Answers were byte-identical across 8 prompts, run to run and alone vs 16 at once
TensorFold 0.6.1 is live.
NVIDIA RTX PRO 6000 Blackwell has landed. The 27B runs directly on it, and NVFP4 checkpoints run in their own 4-bit math, with full-precision mode when quality matters most.
Prompts that arrive together on CUDA now fill in one pass. The 27B with 8 streams runs up to 36% faster, and the slowest first token drops from 0.8 s to 0.1 s.
Flash Next on CUDA: a shared 40,900-token system prompt now answers in under a second instead of 16-56 s. Forks resume from where they split, short requests no longer queue behind long ones, and images work alongside text.
Flash Next prompts fill about 15% faster on a DGX Spark at 8k-32k.
Flash Next on Macs runs up to 10% faster at long context (64k-128k), with the same output.
pip install now works on RTX 40, RTX 50 and RTX PRO cards, with no Docker and no root.
/v1/decisions returns choices, scores and yes/no answers straight from the model.
Native Windows support is experimental.
20 community PRs landed. Thank you all!
Thank you @MiaAI_lab @codengod @shantanugoel @AiMan_993 @edurdias @machinegenie @minviable_org and every contributor and tester.
Special thanks to @plotarmordev @petruspennanen @volatilemarkts for running tests as well!
2
7
32
6,027
Thomas Eastham retweeted
Welcome to Kindle Spark OS!
TL;DR: Add 4GB of RAM to any DGX Spark & faster!
@CK2084, @coffeedev and I figured out how to get 5-10% extra decode and 4GB RAM out of a Spark. We had to replace NVIDIA's OS to do it, but we made it safe and easy.
github.com/kindlingai/kindli…
36
48
409
60,895
Thomas Eastham retweeted
- GLM 5.3 Flash EXL3
- 2x DGX Sparks
- TensorFold
- 4 concurrent prose streams
Pay attention to tok/s and TTFT on each stream.
Buttery smooth experience!
Run GLM 5.3 Flash EXL3 with TensorFold ⚡️
This is a completely new recipe that ushers a whole new level of performance for @NVIDIAAI 2x DGX Sparks.
Conservative default for stability:
- 1M context by default
- 2.7M KV cache pool (!)
- Yes, it's not a typo - 2.7M KV in just two Sparks
- 4 concurrent streams by default
Performance:
- 60 tok/s on prose, single stream.
- 108 tok/s on prose, 4 concurrent streams.
~1,950 prefill tok/s for most context lengths.
Stress-tested to handle a variety of workflows.
This is by far the BEST model to run if you have two DGX Sparks. Extremely smooth experience!
Expect further improvements!
Thanks to @ashxhart for developing such a powerful engine! TensorFold will be used in many of my upcoming recipes.
Get it here:
github.com/MiaAI-Lab/GLM-5.3…
33
27
244
31,972
Spark TP4 no switch... ring cabling
4× GB10 as TP4 vs 2× (2× GB10) as TP2. Same GLM-5.3-Flash EXL3 weights, leased, idle-gated, 500-token replies. @MiaAI_lab recipes
1 stream: TP4 43 tok/s vs TP2 27–29 (+50%)
2 streams on 4 nodes: TP4 65 vs 2×TP2 46–50 (+35%)
4 streams on 4 nodes: TP4 92–99 vs 2×TP2 71–81 (+20%)
Time to first token was under 3 sec. Both TP2 pairs matched each other within 5%.
TL;DR: TP4 wins at every concurrency. 2× TP2 trails by only about 20% under parallel load.
4
1
7
1,901
spark cluster cables just arrived from @digikey tp4 ring (no switch) testing of glm 5.3 flash incoming. running @MiaAI_lab recipe of course. thank you @nvidia for amazing spark hardware networking 🫶
2
3
343
Thomas Eastham retweeted
We spent hours tonight @Alphaschool walking through a day in the life of our kids.
The number of parents getting humbled as they tried to solve workshop challenges or complete academic lessons was impressive.
1
1
23
3,612
This is why an education from @uatx is so relevant today. 👇
BREAKING: Elon Musk on the education he’d recommend for a world of AI in his new interview with CCTV Finance:
“My recommendation for education would be to get as broad-based an education as possible. That includes arts and sciences, engineering, wide general knowledge. Know what to ask the robots for.
You need to be able to formulate the question. And with the broadest possible general knowledge, you'll be better able to formulate the questions, or what they call prompt engineering, I guess. You've got to convey your wishes to the robot.”
52
Thomas Eastham retweeted
Big news!
There are giant gaps in what COULD exist vs what does, in SO many areas of our top universities these days. Education, journalism, sociology, law, etc - are badly ideologically captured and broken.
Excited for @uaustinorg to set higher standards for our civilization.
Today, on Constitution Day, UATX is announcing a new initiative to build the future of legal education, with support from the Federalist Society.
uaustin.org/law
29
42
400
28,064
Thomas Eastham retweeted
I’ve decided it’s time to have a little more fun on X this year! 😄⚡
I get to speak at some amazing events, meet fascinating people, hear great stories, and occasionally find myself in places I never expected to be. So I figured… why not share some of those moments here?
And there’s more!
I’m also excited to launch my new merch. A little Woz spirit, a little fun, and hopefully a few things bring a smile to your face.
This is just the beginning. More adventures, more stories, and more surprises to come!
WozMerch.com
824
978
19,939
2,208,294
Thomas Eastham retweeted
Yeah I see why Dario is afraid
Based on my initial testings DeepSeek v4.1 Flash is outperforming Opus 5 and is very close to Fable 5.1 across the board in frontend web design.
And it's running locally on my little sparks!
260
627
10,811
478,800
Thomas Eastham retweeted
I resigned from nothing today. I’ve spent the last three years doing independent open-source AI R&D, red teaming, and advocacy. The labs are racing straight to self-improving superintelligence and gambling with our lives.
Which is why I’ll spend the next three years doing exactly what I’ve been doing, just at greater scale and speed. And the next. And the next. For as long as it fucking takes.
No more thoughts below.
275
332
6,225
149,865
Thomas Eastham retweeted
If you build your infrastructure right, you shouldn't have to come up with a beautiful prompt every time you use AI
You tell it "yo find me X" and it has enough context to know what you mean
Its why you see me talk to Claude like I'm a 19 yr old broccobro, and still get results
1
1
13
1,359
Thomas Eastham retweeted
Kids are committed to their rock climbing workshop @Alphaschool and went to practice with friends today.
Different levels of success were achieved and one is having a really hard time believing the younger sister might be better at something for once.
2
26
1,678
Thomas Eastham retweeted
MASSIVE performance update to Qwen3.8 Flash Next on a single DGX Spark 🔥
- 46 tok/s for prose, single stream
- 108 tok/s for prose across 4 concurrent streams
- 2,000–2,200 tok/s prefill across all context sizes
- Full numbers in the post below
KV cache is now 1M instead of 1.4M due to stability issues. This was necessary to avoid OOMs.
This is THE best model to run on a single DGX Spark!
Expect further improvements.
Get it here:
github.com/MiaAI-Lab/Qwen3.8…
83
68
777
156,039
