
NVIDIA Flare Architecture: Local LLM Inference for Blackwell and DGX Spark
A technical breakdown of NVIDIA Flare, the open-source local inference engine. Covers Ignition model preparation, Helios Rust serving, CUDA graphs, and agent workloads.
Tag archive

A technical breakdown of NVIDIA Flare, the open-source local inference engine. Covers Ignition model preparation, Helios Rust serving, CUDA graphs, and agent workloads.
Day 20 of my 100-experiment local LLM challenge. Same photo, same prompt, 10 video models — half on a local DGX Spark, half in the cloud. Easy prompt: no difference. Hard prompt: 19x time spread, 11x cost spread, and one failure on each side.
Day 17 of my 100-experiment local LLM challenge. I collected 17 months of my own ChatGPT, Claude, and terminal-CLI history and had a local model (qwen2.5 on a DGX Spark) read it back to me. The point wasn't nostalgia — it was a self-audit: what do I use AI for, how do I ask, and which features am I quietly ignoring?
Day 19 of my 100-experiment local LLM challenge. I got Rusuban-neko (a local qwen2.5:7b) running as the caretaker, took the Slack inbox watcher off the LLM entirely, and parked a read-only Claude on the DGX so I can talk to my secretary from my phone.
![[Day 13] I got a cat to "talk." The biggest wall: the AI couldn't recognize the cat's face](https://media2.dev.to/dynamic/image/width=1000,height=420,fit=cover,gravity=auto,format=auto/https%3A%2F%2Fraw.githubusercontent.com%2FSAETAG%2Fdgx-100-experiments-images%2Fmain%2Fday13-talking-cat-liveportrait%2F01-talking-cat.png)
Day 13 of my 100-experiment local LLM challenge. I drove a single cat image with human facial motion to make a "talking cat" with LivePortrait, instead of putting my own face on camera. It worked — but the hard part wasn't the animation, it was getting the AI to recognize the cat's face. The fix was a no-compile patch for XPose. All local on a DGX Spark.
Day 18 of my 100-experiment local LLM challenge. I built a tiny company of cat-named AI agents to run my task management: Slack as the inbox, Notion as the source of truth, a git-tracked ledger as its memory — and a local qwen on my DGX Spark as the caretaker.
AMD Ryzen AI Halo vs NVIDIA DGX Spark 2026: Which 128GB AI Dev Kit Actually Pays Off
![[Day 12] I tried to build a line-art LoRA from video frames, and the characters' heads fused together](https://media2.dev.to/dynamic/image/width=1000,height=420,fit=cover,gravity=auto,format=auto/https%3A%2F%2Fraw.githubusercontent.com%2FSAETAG%2Fdgx-100-experiments-images%2Fmain%2Fday12-anime-lineart-lora%2F01-before-after.png)
Day 12 of my 100-experiment local LLM challenge. I tried to build a line-art LoRA of my own hand-drawn animation by ripping frames out of the video, instead of preparing stills by hand. Single characters came out broken (fused heads, baked-in backgrounds), I fixed most of it by rebuilding the dataset — but "two characters together" was impossible from video, and the only real fix is to hand-draw the stills. All local on a DGX Spark.
Day 16 of my 100-experiment local LLM challenge. I generated a full Japanese-vocal theme song for my family's cat with ACE-Step 1.5 on a DGX Spark — words, music, and voice, all local. The twist: the audio itself renders in ~10 seconds; the slow part is the AI "thinking" about the song's structure (~8 minutes).
![[Day 11] I turned my cat into anime art — and the AI drew a human girl instead. One photo through IPAdapter pulls it back to a cat](https://media2.dev.to/dynamic/image/width=1000,height=420,fit=cover,gravity=auto,format=auto/https%3A%2F%2Fraw.githubusercontent.com%2FSAETAG%2Fdgx-100-experiments-images%2Fmain%2Fday11-cat-style-transfer%2F02-anime-3way.png)
Day 11 of my 100-experiment local LLM challenge. I tried to restyle my cat into anime, ukiyo-e, oil painting and more — but the AI kept drawing humans. A single reference photo (IPAdapter) drags the result back to being my actual cat, all running locally on a DGX Spark.
![[Day 15] A cat photo became a video — with a meow. Two local video AIs compared: 30s vs 475s](https://media2.dev.to/dynamic/image/width=1000,height=420,fit=cover,gravity=auto,format=auto/https%3A%2F%2Fraw.githubusercontent.com%2FSAETAG%2Fdgx-100-experiments-images%2Fmain%2Fday15-ltx-vs-wan%2F02e-wan-vs-ltx.png)
Day 15 of my 100-experiment local LLM challenge. I turned one photo of my family cat into a short video with two local video AIs on a DGX Spark: LTX-2.3 generated a 4s video WITH a meow sound in 30 seconds; Wan 2.2 took 475 seconds for near-photoreal quality but no audio. Plus a first/last-frame trick that let me design a camera zoom.
![[Day 14] I quizzed an AI on cat meows. It scored worse than random guessing](https://media2.dev.to/dynamic/image/width=1000,height=420,fit=cover,gravity=auto,format=auto/https%3A%2F%2Fraw.githubusercontent.com%2FSAETAG%2Fdgx-100-experiments-images%2Fmain%2Fday14-cat-meow-emotion%2F01e-whisper-gag.png)
Day 14 of my 100-experiment local LLM challenge. I asked a general-purpose audio LLM (Qwen2-Audio) to guess what a cat wants from its meow, using a labeled research dataset of 440 meows. It scored 23.3% on a 3-choice quiz — below random — and a shuffled-options control revealed it wasn't really listening at all. Meanwhile Whisper transcribed the meows as "Thank you for watching!" All local on a DGX Spark.