Leo Laporte (twit.social/@leo)

Podcaster and tech pundit, recovering syndicated radio host, founder and Chief TWiT (long before Elon) at

[Leo Laporte (twit.social/@leo) @leolaporte]

@Tech2Wild Funny you should ask. I just was too impatient with GLM 5.3 Flash, even on dual Sparks. Hermes seemed to be choking it further. DSV4F is so nice (especially with the updates) and the vision version will be even better once the dust settles. Hello deep seek, my old friend.

[Leo Laporte (twit.social/@leo) @leolaporte]

@MikeElgan Nice "subtle" plug for Gastronomad! See you in France in June! 🤩

[Leo Laporte (twit.social/@leo) @leolaporte]

@MParakhin Obviously you're both right, it's simple ... squirrel!

[Leo Laporte (twit.social/@leo) @leolaporte]

@MiaAI_lab I take back everything I said about @MiaAI_lab . Your iteration speed is remarkable. The newest recipe fixes all the bugs I found and I'm back to EXL3 with Hermes on dual sparks. And I love the one line switch between abliterated and stock. Still the Queen!

[Leo Laporte (twit.social/@leo) @leolaporte]

@MichaelGannotti @MiaAI_lab @NousResearch @sfxnz I really wanted to like the speed of EXL3 but I kept getting empty answers. NVFP4 won on reliability, not speed.

[Leo Laporte (twit.social/@leo) @leolaporte]

No disrepect to @MiaAI_lab and @MichaelGannotti but after benchmarking all the recipes I could find for GLM 5.3-Flash with @NousResearch's Hermes on dual Sparks the one that best fits my use case is @sfxnz's open vLLM recipe. I'm not tokenmaxxing I just want a smart reliable local model.

[Leo Laporte (twit.social/@leo) @leolaporte]

@witcheer Bless you @NousResearch - Hermes is heaven and your token plan is the heavenly choir singing to my local AI.

[Leo Laporte (twit.social/@leo) @leolaporte]

@ShawnTempesta Just to put this all in perspective, when I started in radio -- 50 years ago! -- my first paying gig was $700 a month. I didn't own a car. I didn't own a TV. But man, I loved that job.

[Leo Laporte (twit.social/@leo) @leolaporte]

@film_girl @openclaw @github @steipete @BradGroux @joshavant Love me some claw. By the way, you were great yesterday on Intelligent Machines. Thanks for hanging in there for all two hours!

[Leo Laporte (twit.social/@leo) @leolaporte]

@Teknium @witcheer Yes please! Count me in!

[Leo Laporte (twit.social/@leo) @leolaporte]

What kind of local models could you run on an M5 Ultra with 512GB of unified RAM at 1.2TB/sec? And how much would you be willing to pay for that? It's got to be $20K.

But where are they getting the memory??

[Leo Laporte (twit.social/@leo) @leolaporte]

Interesting. I asked Fable to design seven of the hardest coding problems it could and gave them to Ox Alpha, Grok 4.6, and GLM 5.3. Ox Alpha and Grok both aced the tests. GLM 5.3 scored 73%.

I don't know what Ox Alpha is but if it ends up being open weights this will be a watershed moment.

[Leo Laporte (twit.social/@leo) @leolaporte]

I'd love to see it - I run your recipe for DSV4F on my dual sparks and it's a thing of beauty. Thank you!

I'm running Orca's Qwen3.8-27B-Uncensored-Q4_K_M.gguf on my 3090. Quant is Q4_K_M, KV cache is unquantized F16, and effective context is 24,576 tokens (49,152 split across 2 parallel slots).

[Leo Laporte (twit.social/@leo) @leolaporte]

@film_girl oops. I meant GGUF Q4_K_M - the RTX doesn't do FP4 :-(.

[Leo Laporte (twit.social/@leo) @leolaporte]

On a whim I bought a used RTX 3090 last week and upgraded my old gaming machine. Qwen3.8 27B NVFP4 runs beautifully on it. And with the new vision capabilities I've retired the old Qwen VL on the Framework.

3.8 is so good at vision, it even sees me as a young man. Well, younger. 50s-60s. I'm keeping it!

[Leo Laporte (twit.social/@leo) @leolaporte]

I was questioning my common sense when I traded my 3070 in for a used RTX 3090. And then I knew I'd gone totally crazy when I bought *two* DGX Sparks (spending my kids inheritance).

And yet, today, running Hermes with Qwen 3.8 27B on one machine and DeepSeek V4 Flash 731 on the Sparks, all totally local, I think maybe I wasn't so nuts after all.

We live in amazing times.

[Leo Laporte (twit.social/@leo) @leolaporte]

If there were a doomsday clock for a world built on software, I would set it to one second to midnight.

[Leo Laporte (twit.social/@leo) @leolaporte]

I’m so glad we were at black hat for this. Literally the most important story in tech today. We are at an inflection point. We knew this would happen eventually but it’s here now and it’s both exhilarating and terrifying

https://t.co/J2VXDMrubM

[Leo Laporte (twit.social/@leo) @leolaporte]

@Teknium It’s cheap until you decide to buy dual Sparks

[Leo Laporte (twit.social/@leo) @leolaporte]

@SGgrc @richcampbell and I are enjoying Threatlocker branded Guinesses at the Rí Rá Irish Pub ☘️ at Black Hat! #sponsor Sláinte!