iiinova — where your digital self keeps evolving

Shawn Reed @shekari.bsky.social · 04/10/2026

It started as "how hard can it be?" One person, a good model, a good coding assistant and a lot of patience got it here. I hope it encourages you to grab your favorite LLM and framework and build something from scratch. #StrixHalo #LocalLLM #AMD #ROCm

010

Shawn Reed @shekari.bsky.social · 04/10/2026

The model is Qwen3.8-Flash-Next, a ~180B MoE, converted for the engine - a ~115 GiB download instead of the 360 GB checkpoint: huggingface.co/wemoh/Qwen3.8-Flash-Next-strixw

huggingface.co

wemoh/Qwen3.8-Flash-Next-strixw · Hugging Face

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

000

Shawn Reed @shekari.bsky.social · 04/10/2026

The speed holds at depth because the model's sparse attention never looks at more than ~2k keys per token, and the engine's own multi-token prediction drafts up to 5 tokens per step. Measured in real agentic coding sessions, not a synthetic benchmark.

000

Shawn Reed @shekari.bsky.social · 04/10/2026

A big thank you to Donato Capitella for all his dedication to the Strix Halo community - his toolboxes, benchmarks and terminal-bench-mini, shared openly, were a constant source of inspiration and kept my numbers honest. strix-halo-toolboxes.com

000

Shawn Reed @shekari.bsky.social · 04/10/2026

Heads-up: the weights are still uploading to Hugging Face from my home connection - they'll be complete around 7:30 PM Pacific today. The code and the Quick start are up now; the download step works once the upload finishes.

000

Shawn Reed @shekari.bsky.social · 04/10/2026

Back in March I was excited that Qwen's open models could do real coding locally. So I went further: I wrote my own inference engine for them, from scratch, for AMD Strix Halo. ~47-50 tokens/s, flat out to 492k tokens of context, on one mini PC. github.com/shawnshekari/strixite #StrixHalo #LocalLLM

strixite banner: a crystalline owl perched on a computer chip, with the text "strixite - LLM inference, from scratch, for AMD Strix Halo".

500

Shawn Reed @shekari.bsky.social · 28/08/2026

Those are the exact correct steps to becoming a soldering expert.

100

Shawn Reed @shekari.bsky.social · 15/08/2026

github.com/shawnshekari... github.com/shawnshekari... Latest including stream mode is not pushed to the repo.

000

Shawn Reed @shekari.bsky.social · 15/08/2026

Spent some time with claude optimizing mandelbrot rendering on my RC2014 Pro Z80. Down to 24 seconds! Also showing off a new pixel mode in mt Bridge application. A pyGame terminal emulator that also exposes an MCP so your model can take over control of your retro computer. youtu.be/uB60Z52KaVk

100

Shawn Reed @shekari.bsky.social · 27/03/2026

Chad Bianco should arrest Chad Bianco.

000

Shawn Reed @shekari.bsky.social · 14/03/2026

Opencode and qwen3.5-122b-a10b local

000

Shawn Reed @shekari.bsky.social · 12/03/2026

Sorry but from that article “The consumed water is generally the water that evaporates or is otherwise taken out of immediate human usage. “ that is not consumed, it is a state change from liquid to gas but it will change back to liquid. Nothing is lost or consumed

000

Shawn Reed @shekari.bsky.social · 12/03/2026

They don’t consume water nor do they consume fossil fuels. They consume electricity and produce heat. The water is utilized for cooling.

100

Shawn Reed @shekari.bsky.social · 08/03/2026

Qwen's newish open-weights models are very good. I was able to get a couple of them 27B (dense) and 122B (MoE) to get real coding done yesterday. So local coding with the only expense being the the power used. It is not fast but it is still 10X what I could do in a day. #llama.cpp #Fedora #ROCm #AMD

130

Shawn Reed @shekari.bsky.social · 25/02/2026

Love Anna Weatherup's new album! I was so excited to get the CDs delivered from down under and signed to boot. Thank you Anna and especially for sharing your music. annaweatherup.com www.instagram.com/annaweatheru... www.youtube.com/annaweatherup

010

Shawn Reed @shekari.bsky.social · 11/02/2026

Common Saints - Cinema 3000 Afternoon stress relief with a signed CD. commonsaints.com

010

Shawn Reed @shekari.bsky.social · 22/01/2026

The cameras should be banned. The data will always be misused and that risk outweighs any positive use.

000

Shawn Reed @shekari.bsky.social · 21/01/2026

Now playing, first CDs purchased in many years and signed even. Thank you Maz for the wonderful music. mazoconnor.com

000

Shawn Reed @shekari.bsky.social · 16/01/2026

Note the 10" powered sub is not shown but is obviously a key component. #separates #hifi #greaterthanthesumofparts

000

Shawn Reed @shekari.bsky.social · 16/01/2026

Initial test of my streaming replacement system is going very well. I think I can tough it out. So long Spotify, Google, and Tidal I got it from here. #Fosi #klipsch #audiolab

Fosi audio and Klipse speakers and Audiolab CD transport

210

Shawn Reed @shekari.bsky.social · 04/01/2026

I am quitting music streaming and I am quite excited about it too. #RedBookAudio #PCM #44khz #16bit

010

Shawn Reed @shekari.bsky.social · 01/01/2026

Passwords in the cloud? Pass

000

Shawn Reed @shekari.bsky.social · 19/12/2025

If anyone still has it installed any longer.

000

Shawn Reed @shekari.bsky.social · 17/12/2025

So long Firefox, it was a good run.

010

Shawn Reed @shekari.bsky.social · 10/12/2025

The utility to use is called Linux.

000

Shawn Reed @shekari.bsky.social · 03/12/2025

Outsource my data storage? Pass.

000

Shawn Reed @shekari.bsky.social · 23/11/2025

Sounds like anthropomorphism to me. Is the model being devious or just taking the most efficient path to solve the problem.

000

Shawn Reed @shekari.bsky.social · 04/11/2025

Whoa!! Seeing some insane improvements with #ROCm 7.1! #AMD VAE decoding is lightning fast.

010

Shawn Reed @shekari.bsky.social · 03/11/2025

I think cooling towers are typically evaporative

000

Shawn Reed @shekari.bsky.social · 03/11/2025

Also contaminated water through cooling towers means contaminated air

100

Shawn Reed @shekari.bsky.social · 03/11/2025

Cost. How do you get it where you need it.

000

Shawn Reed @shekari.bsky.social · 27/10/2025

Start with an rc2014 and build it first.

000

Shawn Reed @shekari.bsky.social · 25/10/2025

+1 on shell preference as well.

010

Shawn Reed @shekari.bsky.social · 23/10/2025

#AMD 's #ROCm continues to improve! 43 tokens/sec with a 30B MOE Q8 and 512k of context < $2k mini PC For those worried about power consumption it just means I need to run the heater less now.

010

Reposted by Shawn Reed

François Zaninotto @francoisz.bsky.social · 21/10/2025

Agentic Search surpasses Retrieval Augmented Generation (RAG): Teaching an agent to use grep is far more efficient than chunking, embedding and similarity search. #AI www.nicolasbustamante.com/p/the-rag-ob...

nicolasbustamante.com

The RAG Obituary: Killed by Agents, Buried by Context Windows

Why Retrieval-Augmented Generation Won’t Survive the Context Revolution and the End of Chunking, Embeddings, and Rerankers as We Know Them.

031

Shawn Reed @shekari.bsky.social · 16/10/2025

No thanks Grok-Code-Fast-1 I'll take it from here.

010

Shawn Reed @shekari.bsky.social · 15/10/2025

I literally woke up and thought I can't wait to get back to testing multi-modal embeddings today. #Memgraph #MCP #llama.cpp

010