Freedom Engineering

Build systems. Reduce dependency. Increase freedom.

[Freedom Engineering @FreedEngineer]

The blazux/qwen3.8-Flash-DGX recipe memory-maps that table read-only from NVMe and fetches only the required rows per token, keeping it out of shared memory.
https://t.co/GRlgUfIFEv

[Freedom Engineering @FreedEngineer]

Running RadixArk/Qwen3.8-Flash-Next-NVFP4 on one 128GB DGX Spark with vLLM: TP=1, MTP=2, 262K context, one concurrent sequence.
It uses roughly 97 GiB of unified memory , including ~79.4 GiB model weights and ~14.3 GiB KV cache.
Continuous coding reaches 30+ tokens/s.

[Freedom Engineering @FreedEngineer]

The checkpoint occupies ~126 GiB on disk, plus ~19 GiB for the Docker/vLLM image: approximately 145 GiB total.
The checkpoint includes a ~47.7 GiB FP8 n-gram/PLE table.

[Freedom Engineering @FreedEngineer]

@jun_song No you need your sleep, don't fight the nature

[Freedom Engineering @FreedEngineer]

And the usage is separated from cursor's other usage which is great, it's a great addtion for my work stack, the more the merrier.

[Freedom Engineering @FreedEngineer]

Grok bot is really good, it's has the iOS app, so I can test it when I am on the road.

[Freedom Engineering @FreedEngineer]

And the iOS app is in sync with the Mac app, which is awesome feature I didn't get from others.

[Freedom Engineering @FreedEngineer]

@TheAhmadOsman crazy world, 0731 didn’t last 30 days

[Freedom Engineering @FreedEngineer]

@Tech2Wild lol

[Freedom Engineering @FreedEngineer]

@NeoAIForecast lol

[Freedom Engineering @FreedEngineer]

Apple’s M5 Ultra and NVIDIA’s DGX Spark are worth comparing, but chasing the newest AI hardware is a loser’s game. Use your current setup until it hits a real bottleneck. Upgrade only when the gain justifies the cost. I bought DGX Spark to learn—and it’s doing exactly that.

[Freedom Engineering @FreedEngineer]

@shiiba__ You can just have a Codex set up it for you, no need following instructions

[Freedom Engineering @FreedEngineer]

Each model has its own use case, with experiments on DGX Spark I understand it more

[Freedom Engineering @FreedEngineer]

@yacineMTB yes, and now what I do is trying to use it to help me in a lot of ways. the limiation is on me not ai model.

[kache @yacineMTB]

100% correct, deepseek flash is better than most human beings

[Freedom Engineering @FreedEngineer]

@Tech2Wild It will be great if it's optimized for one DGX spark.

[Freedom Engineering @FreedEngineer]

@MiaAI_lab Might change soon😜

[Freedom Engineering @FreedEngineer]

12 days ago, when I bought DGX Spark it's was 5100 USD tax included, now, 12 days later, the price is just hiked 450 USD. Crazy world.

[Freedom Engineering @FreedEngineer]

@InvestingAddict Insane. I need add 2000 USD extra for a DGX spark, the sauna seems cheap.

[Freedom Engineering @FreedEngineer]

@OskarWieckowicz You forget the skills you don't need.