Joe Muller

howtospark . com is live! I'm treating it as a living notebook for my Spark experiments and have a bunch of things I want to add that will help Spark owners hit the ground running

@hnakamur2 Ah let me update that...my bad haha

@BosonJoe In the "Put the deployment config on both nodes" section, it says: # on BOTH nodes git clone github.com/jtmuller5/howt… ~/howtospark

The DeepSeek V4 Flash Recipe has been updated 🧑‍🍳 There was a bug in the spec-decode config Bumping k from 3->5 increased code content decode tok/sec from 62.5 → 73.2 tok/s (+17%) Working on updating to vLLM 0.26.0 next...

@plotarmordev For code content, everything else stayed at about 42-45 tok/sec

@BosonJoe 17% just from k 3 to 5?

@hnakamur2 Hi, they should all be on the website not the GitHub

@BosonJoe Hi, thanks for sharing your recipe! It looks like the link to github.com/jtmuller5/howt… is broken and returns a 404 error.

@danushman I can empathize with this

made the mistake of trying to use Grok for something... what a POS. Come back in 9 hours, and it's stuck, for the stupidest possible reason, got nothing done. Claude is taking over. Never again. Grok 4.5 is trash and I don't care what anyone says. Absolute trash.

16 hours people

@MrPeterLMorris NVFP4 is almost lossless for this model. You could also fit the full BF 16 variant if you wanted to it would just be slower

@BosonJoe I would need to know how intelligence compares. I've seen tricks to speed up models that made them useless.

Qwen3.6 27B NVFP4 Single Spark Recipe ⚡️ 🚀 vLLM 0.26.0, no patch 🪣 262K context 👾 31 tok/sec decode You could probably run this and the 35B MoE Qwen model on the same Spark if you wanted the best of both howtospark.com/recipes/qwen3-…

@adisingh NDAs are a cancer

I've reached that point of my company where I hand out NDA's like free candy

@kkiran Personally I would choose moe models. Anything dense is going to be pretty slow

@BosonJoe contemplated both but Claude says this for a single Spark - 35B and nemo3. If they don't play well, then 27B + 35B or Nemo 3. Thank for the amazing work!

@basewtf Yes theres room

@BosonJoe Could run 27B and 35B concurrently on 1 spark?

cc @LannyTheClever

I still use Claude for most things, especially running benchmark tests on the Sparks (it's easiest because I don't have to use memory to host the model I'm using to run experiments) When I do use local models I use Deepseek V4 Flash and/or Qwen 3.6 35B A3B. I only do code related work

@BosonJoe @NVIDIAAI It seems like you’re still digging into the GLM 5.2. I’d be curious what you end up using as a default and what you feel like they’re best used for. I’m wondering as well if Qwen 3.8 open weights drop and the 35B of that ends up being a major jump forward for individual sparks

GLM 5.2 on 2 @NVIDIAAI DGX Sparks @ 24.7 tok/sec 🚀 Next step is to claw back some quality while keeping the speed steady We've come a long way. When I stared it was 4 tok/sec 🐌

@vectorpolygon @CtrlAltDwayne They made inference faster for every topic by making the model less capable at biology and cyber security Do you know how model inference works?

@BosonJoe @CtrlAltDwayne Use your intelligence.

It's insane how token efficient Opus 5 is. I've been hammering it and my weekly usage has barely moved. How did Anthropic make such a capable model so token efficient without sacrificing intelligence and taste?

@vectorpolygon @CtrlAltDwayne That makes no sense

@CtrlAltDwayne By sacrificing intelligence that matters to people who actually need it, while making people like you say "it's insane".

@imost @RyanFogleOP12 Dang, looking into it. Thanks for the heads up, I should have a fix out this morning

@BosonJoe @RyanFogleOP12 Just tried to submit a benchmark for your GLM 5.2 recipe but got this. Let me know if there's anything I need to change on my side.

First contributor to howtospark! Thanks @RyanFogleOP12 Check out the MCP server to join up 🤝

@gnubyte_ @NVIDIAAI No, I would classify this work as experimental/educational Best model on 2 sparks is still Deepseek v4 flash

@BosonJoe @NVIDIAAI Hey man so have you been able to build anything with it yet? It seems like to me the best money has just been on the Qwen3.6-35B NVFP4 model that sparks sorta ship with or try to default to.

@nietonchique @NVIDIAAI Not sped up, but this is very experimental and I would not use it to do anything yet

@BosonJoe @NVIDIAAI speeduped video, lier

@SteveSTechMan @NVIDIAAI howtospark.com/recipes/glm-5-…

@BosonJoe @NVIDIAAI What’s the recipe?