Mike Hutu

AI Native Dev | Open-source hunter | Focused on building with- and on top of foundation & local models | Keeping up with the AI era |

[Elon Musk @elonmusk]

Grok Build is now open source

[Mike Hutu @MichaelHutu]

I don't know about 3 times the size, but yes naturally the size should grow the more data it has.

But if you look at what they did with this 54Gb model. What stops them for repeating the same with a bigger model, after they get feedback and take the method even further? Afterall, this was the first batch.

Cutting out 90% of the size and keeping that much of the quality is unbelievable fr. If it grows to 100Gb or even 150Gb, it would still fit in to a consumer grade GPU.

[Mike Hutu @MichaelHutu]

Can you create realistic high quality images fully local with DGX SPARK? - YES YOU CAN👁️

I've been testing out the creative side of @NVIDIAAI DGX-Spark.

It often surprises me how much you can do with this tiny little machine from your dark basement, fully local.

This one is Ideogram 4, and i took it on a ride with a single DGX-Spark.

I wanted to test if it could recreate images, gave an image (left image) to Gemma4 model to extract all the details and create a json-prompt to recreate the image. Took the prompt to Ideogram workflow, waited 2 minutes cold to get crazy good quality of image.

The small details in the image are unbelievable, the fabrics, the little hair on her chin and arm, even the pores of her skin. This was only a 10min job from setting up the workflow, find an image and test fast without even checking the json-prompt, just copy paste.

Max load was around 35Gb/128Gb.

With a single DGX-Spark, you can run this workflow, Gemma4/Qwen3.6 35b a3b (or both), Embedding and reranking models, orchestrated by Hermes or any harness that you can connect custom Openai compatible endpoint.

[Mike Hutu @MichaelHutu]

The strongest western open weight model has been published.

Meet Inkling the strongest open weights model from the westworld.

975B parameter model made by @thinkymachines 👀

@UnslothAI said ”hold my beer” before my feed updated and the next thing i saw was a 1-bit GGUF model that fits in 280Gb

It means that now
3x @NVIDIAAI DGX-Spark has one more, and i need one more 🫣 actually two more..

[Mike Hutu @MichaelHutu]

@MiaAI_lab Made some plans for your weekend:
https://t.co/THiaWkjK5U

[Google Gemma @googlegemma]

We’re rolling out some big improvements to Gemma 4, fueled by incredible community feedback and contributions!

Here is a breakdown of what’s being fixed and updated in this release: 🧵👇

[Mike Hutu @MichaelHutu]

Dear @deepseek_ai @Zai_org @MiniMax_AI @Alibaba_Qwen @XiaomiMiMo @Kimi_Moonshot

Could you please release lightweight open source flash variants of your future frontier models?

Thank you!

[Mia @MiaAI_lab]

Dear @deepseek_ai @Zai_org @MiniMax_AI @Alibaba_Qwen @XiaomiMiMo @Kimi_Moonshot

Could you please release lightweight open source Flash variants of your future frontier models?

Thank you!

[Mike Hutu @MichaelHutu]

@praiseneh @PrismML You can run 5-10 or even more agents with one model with that gpu 👀 each with own system prompt, loops n tools using @vllm_project

[Mike Hutu @MichaelHutu]

@mr_r0b0t @OssamaEl56098 @NousResearch 🦍

[Mike Hutu @MichaelHutu]

@mr_r0b0t @OssamaEl56098 @NousResearch This feels like early 2000s and finding a new cheat code ⬅️⬆️➡️⬇️⏺️⏺️⏹️⏹️⬆️⬅️➡️⬇️

[Mike Hutu @MichaelHutu]

@MiaAI_lab Missing 2… or 4.. Just saw someones 6x tower and got FoMo. I need to open up a OF or something 🫣

[Mike Hutu @MichaelHutu]

@mr_r0b0t @OssamaEl56098 @NousResearch Love your thinking game 🔥

[Mike Hutu @MichaelHutu]

We are gonna have a GREAT Q4 for Local AI.
So excited about everything.
The 1 to 4-bit quality is getting seriously good.

If you're still thinking if you should invest $4k to get a local server, you got to take action.

@NVIDIAAI DGX-Spark is still sold for ~$4k, but people are buying them like there's no tomorrow and most of the buyers are probably just individuals that have been taking notes.

Check what's happening in the memory market and prices for the last 6months and calculate yourself what's the trend for 2027. Those companies have basically sold everything they can produce up to 2028 and they are actively building new factories.

I already see many companies that are hiring Local AI engineers. When one achieves in it, other will repeat the same.. Eventually it will become as popular as Claude right now amongst enterprises.

Data is becoming the only valuable resource that can't be artificially made using AI. Everything else is now possible to replicate or better said reverse engineered.

Own your AI -> Own your DATA.

[Tencent Hy @TencentHunyuan]

We’ve just released the 1-bit & 4-bit version of Hy3, a flagship-scale 295B model that can be served on a single GPU. 👌

Run Hy3 with llama.cpp, enable MTP, and experience powerful intelligence on dramatically lower hardware.🚀🚀🚀

Can’t wait to see what you build.

#Hy3 #Hy #GGUF #llamacpp

[Mike Hutu @MichaelHutu]

@MiaAI_lab It will.. they already proved their competence. If you think about it, it’s july - the most low traffic month of the year and we see people/teams shipping every week. Maybe they have time from their initial job or we’ll see an explosive Q4

[Wësche @WescheNex1q]

MiMo V2.5 scored TrueScore 84.7 on our 64-scenario DGX Spark benchmark

#2 overall, edging out Nemotron Ultra 550B (82.7) on half the hardware.

2× Spark, TP2, NVFP4.

Perfect 100s in structured output, classification & efficiency. 97.8 visual, 88.1 code, 79.4 agentic.

It generated these from single prompts 👆

[Mike Hutu @MichaelHutu]

@anuntrapid_auto Thats a nice $30k server you got there 👀 how do you handle the heat?

[Mike Hutu @MichaelHutu]

@MiaAI_lab Exactly, my iphone is already occupied by social media and telegram 😅

This is unbelievable fr.

[Hugging Models @HuggingModels]

Big news for AI on a budget. GLM 5.2 Colibri int4 is a Mixture of Experts model that runs entirely on your CPU. No GPU needed. It's fast, efficient, and opens up advanced language capabilities to anyone with a standard computer. This is a game changer for offline AI.