Tech accelerationist🛫 You don't know what's coming, but you'd better hang on tight
Asia
Joined January 2026
- Tweets672
- Following259
- Followers88
- Likes2K
Pinned Tweet
Anthropic took Fable 5 away from you. If that made you nervous about cloud AI dependency, you're not alone. Here's how to actually run AI locally, with realistic expectations:
(link's in the thread)
Made this demo tonight on a mini-PC in the spare room. No cloud, no subscription, no API - MiniMax H3, fully local, lip sync & all.
The future doesn't wait for permission.
The how: MiniMax H3 (Comfy-Org pruned int8 build) in ComfyUI on an AMD Strix Halo mini-PC, 128GB unified memory. 864x480 @24fps, ~14 min render. Voice, lip sync, SFX and picture in one pass. Zero bytes left the building.
Rellen Kray retweeted
I ran some basic benchmarks to compare all of the top Qwen 3.8 Flash models on Halo Strix.
This model is amazing, the first time I've felt like I have real frontier intelligence at home.
The Ciru-Orca model is probably my favorite of the ones below.
evidence(for agents): github.com/ciru-ai/strix-sho…
benchmark page(for humans):
llm.ciru.ai/research/strix-s…
The AMD bet I made earlier this year is definitely paying off👍🏻
Ryzen SoC Linux users buckle up! @AMD Boosting AI/LLM Performance For @AMDRyzen @AMDRadeon iGPUs As Much As 18~23% With Linux 7.4
AMD PerfOpt benchmarks across Strix Halo (@FrameworkPuter Desktop), Strix Point, and Krackan Point hardware.
phoronix.com/review/amd-perf…
Rellen Kray retweeted
Insane Gufo performance at Strix Halo - prompt processing 1191.35 tok/s, token generation 36.13 tok/s at 40% of 260k context used by AI agents.
Keeping an eye on this one. My ISP already hates me for all the Qwen-downloading I've done in the past few weeks.
Run Qwen3.8-Flash-Next on one AMD Strix Halo (128 GB, Linux).
One 78 GB file, full 262K context, tested at every depth.
Reading 301-870 tok/s, writing 17.5-46.7 tok/s.
Copy-paste setup, tested on a clean Fedora + Ubuntu:
huggingface.co/PaoAI/Qwen3.8…
Engine: @ilintar's ROCm branch
llama.cpp PR 29030 swaps mmap for direct tensor reads. Strix Halo prompt processing: +121% on pp512, +75% on Qwen3.8-Flash. Unified memory was stalling the iGPU. Same answers, less waiting. If you run local on an AMD APU, this might be the patch🤞
llama.cpp PR #29030 by @ilintar: direct reads replace mmap for lazy tensor rows. All passed! Need reviewer!
-Strix Halo pp512: +121% 🚀
-Qwen3.8-Flash pp2048: +75%
Same output, way faster. Huge for iGPUs and DGX Spark where mmap stalls on unified memory.
github.com/ggml-org/llama.cp…
Tested on my Strix Halo (gfx1151, 128GB unified) with Qwen-Image-2.1 INT8 today. 2048x2048 went from 33 min on vanilla ComfyUI attention to 154 sec🤯 1024 drafts in ~35s.
This is their own teapot reference prompt, straight off my box. Card text rendered, not retouched.
Qwen 2.1 Image is amazing.
I made a comfyui custom node that improves performance on AMD Strix by over 700%.
Ciru AMD Halo Qwen 2.1 Turbo
From over 20 minutes for 2k image to under 200 seconds.
Adjustable quality trade offs, with ciru total your 1024 30 step image goes from 64 seconds to as fast as 25
2048 image in this tweet goes from 45 minutes to 183 seconds at default Ciru Turbo settings.
My first image optimization, please give feedback!
Use: comfy node install ciru-image-accelerator
or
github.com/ciru-ai/ComfyUI-C…
Halogen 0.1.4 on Strix Halo: one podman command, ROCm in-container, host Vulkan untouched, 256K ctx. Cleanest gfx1151 engine install yet👍
@high52weeks
Halogen + Qwen3.8-27B on Strix Halo: 607 t/s prefill (2.3x the tuned llama.cpp fork I was using before), 49.6 t/s code decode - even beat the 44.2 claim.
Bonus find: the bench harness @BenchLocalAI has a hidden 300s timeout silently failing slow engines. Half our old CLI-40 scores measured a timer, not the model.
I moved into AMD/Strix Halo hardware for very similar reasons, so I've also been glad to see some serious optimization improvements there.
As local inference hardware goes, people generally target (in order): NVIDIA GPUs, Apple stuff, DGX Spark, and then the Strix Halo. I got the Framework because it was a bit more cost-effective than the DGX Spark, but I also just like the attitude of the company and wanted to invest in a broader hardware ecosystem. The Framework also has some hardware differences that make it attractive as a general-purpose Linux machine. It's x86-64 rather than ARM, and exposes PCIe which allows me to hook up a discrete external GPU. The Spark, to my knowledge, does not allow that. More on eGPU another time.
Rellen Kray retweeted
Proud to see Votari featured in CoinGeek, sharing what we’ve been building and where we’re heading.
The idea behind Votari is simple: you should be able to check the evidence behind an election result while protecting voters’ privacy.
That principle has driven a lot of work behind the scenes. Votari is now available on iOS and Android, with different participation options to suit different elections.
Our next chapter is working with associations, universities, communities and other organisations to put that work into practice. Each deployment is an opportunity to learn, improve and demonstrate what independently verifiable voting can offer.
Thank you to everyone who has tested the app, challenged our thinking and helped us get this far.
If your organisation runs elections and would like to explore a pilot, I’d love to hear from you.
Here’s the full story:
👀
Recently, we worked on something special: pushing ROCm to beat Vulkan on Strix Halo.
Lucebox engine (ROCm) and Llama.cpp (Vulkan) ran on the same Strix Halo on the same day with the same prompt. With the DSpark drafter Lucebox ROCm decodes at 41.9 tok/s at 8K and 38.0 tok/s at 123K prompt tokens.
That is 28% to 47% faster than llama.cpp and Vulkan v0.7.5 running the same drafter and its prefill is 17% to 47% faster 🧵
It's 2026 and I'm downloading a ~100GB file in one hit... and there are no MPAA lawyers knocking on my door
Got my strix halo running @UnslothAI's Qwen3.6-35B-MTP-UD-Q4_K_M with halo-box/strix-llama.cpp:
~78-81 t/s; scored a healthy 83-88/100 on @BenchLocalAI's HermesAgent-20 test. I am happy.
github.com/halo-box/strix-ll…
