Mike Key
@1337hero
Senior Software Engineer going hard on AI.
Building local LLMs & AI agents
Homelab - 3x AMD R9700 & Strix Halo
Shipping code @
279 Following 2.2K Followers
I'd own a spark in a heart beat if it came with 1.2TB/s of memory bandwidth.
If you're considering a Spark follow people like @mr_r0b0t, @aijoey and @Tech2Wild there are others I'm forgetting for sure.
They're giving the most honest feedback regarding them.
From my perspective if you can afford Discrete GPUs always go Discrete GPU. 3090 and up if you can.
It's really that simple. There's a reason gpu's have gone up significantly during this time and continue to have price increases along with stock decreases.
Nvidia also isn't giving out free 5090's and pro 6000's to creators to review or sell them for a reason.
As long as you do your research you'll be happy either way.
Show more
The difference is:
DHH isn’t out to make Linux political. He’s out to make it apolitical. And that’s really pissing off the entrenched Linux clowns.
Also he and the people working with him are generally a) way better programmers, b) better people and c) adjusted individuals with families, friends and a life full of joy.
Show more
i wanted a better understanding of historical political boundaries so i made this thing where it shows you them on a timeline, but also overlays selectable historical maps at that time over the corresponding region
Show more
REAP when done by competent people
I love Tailscale + Headscale
I can be anywhere in the world with Starlink and hit my home machine and then with my 3 R9700's spin up a local model and start working on a PR.
We truly live in remarkable times.
Show more
Some spots really do look just like the photos you see online.
Day 3 of my Oregon road trip.
Why is Github just utter garbage anymore.
Microsoft hiring practices?
Github employees vibe coding?
Or is AI cover for hiring?
I have been working hard on ROCmFPX the last two days and now you can run NVFP4 natively inside ROCmFPX on AMD Hardware. You no longer need to convert it, just download and go. MTP works too.
This is a pretty big update because in the past you needed to do a conversion and you can still do conversions (only if you want to) from NVFP4 - ROCmFP4. That's also been made better.
Next updates I am working on are better ROCmFP2 and ROCmFP3.
My goal is simple. Make ROCmFPX work on any and all AMD hardware whether you run Windows or Linux. It's optimized for ALL AMD Hardware. Well, RDNA 1 through RDNA 4 etc. And yes, this in theory should run on your Asus Rog Ally's too. Z1 and Z2.
Show more
Managed to bump the generation speed with a minor hit on prefill with finding out some things were wrong with my llama-swap config and my hardware config - editing grub enabled some things that should speed up all my runs on everything. Time to re-bench everything.
Show more
3090's havin' a day today
For the doubters, Opus 4.6 at home on your own hardware.
The 15~2-% materialized on agentic coding, terminal/computer use and office/workflow tasks making it one of the best (if not the best) dense open models for local agentic work in it's size.
I received a bunch comments saying this would never be this good. But here we are.
Show more
My Qwen3.8-27B predictions:
- +15-20% intelligence jump
- MTP support landing in llama.cpp
- Best dense model for 24GB VRAM
- Will rule local agentic workflows
- Close enough to Opus 4.8 / V4-Flash that many dump subs
- Resets “good enough to self-host” for 2026
what's yours?
Show more
Time to get Qwen3.8-27B up on LocalMaxxing
Qwen updated the HF model card w/ some interesting details about 3.8:
- Native vision: images + hour-long videos (this one surprises me the most)
- support for popular agentic harnesses & coding tools
- thinking is toggleable per request
- supports reasoning_effort (xhigh default) - supports preserve_thinking
• 262K native ctx expandable to 1M ctx
Funny; they call it a casual model and a lot mentions about support for popular harnesses and development tools, making it easier to integrate into your existing stack.
This could be the best local model for Hermes and OpenClaw and such...
TBD....
Show more
This is your reminder to take a break from your computer every once and awhile.
Good for the soul.
Good for your creativity.
Exceptionally good for your health.
Vibe coding can wait. Or just leave your agent to it, go for a hike; get a status update when you get home.🤘
Show more
Do me a favor.
Ask Qwen3.6-27B what its knowledge cut off date is and post the response below.
It told me it's 2024. Surprising for a model released this year.
Figured it out eventually
Muse is the slowest dense model I've ever attempted to run at Q8_0 and I don't know what the problem is. All day and I can't get it faster than 16tok/s / 661.7 ppt - pretty much WORSE than me trying to run DSv4 1 bit.
Show more
Muse is the slowest dense model I've ever attempted to run at Q8_0 and I don't know what the problem is. All day and I can't get it faster than 16tok/s / 661.7 ppt - pretty much WORSE than me trying to run DSv4 1 bit.
Show more
