Musician & producer 🎛️
Port Stephens, Australia
Joined January 2025
- Tweets643
- Following154
- Followers207
- Likes7
Pinned Tweet
New #axonis drum and bass: "Wait for This" 🎧
172 BPM, whispered vocals climbing into full belts, rolling reese bass and dark neon rain. The video was generated and edited entirely on our own servers here in Australia, no cloud.
"I can wait for this, I can wait and I will never miss."
#DnB #DrumAndBass #AIVideo #LocalAI
Signal in the Static 📡 New track and music video. Every shot was rendered on my own local AI hardware and cut to the beat. 🎶 #AIVideo
Just released Axonis GPU Monitor v1.2.1, a free Windows app that shows your local AI servers (llama.cpp, vLLM, ComfyUI) and NVIDIA GPUs at a glance. Built by my local AI developer team, running on my own hardware.
Download: github.com/axonis22/axonis-g…
@ggerganov @vllm_project @ComfyUI
#LocalLLM #LocalLLaMA #llamacpp #OpenSource
Qwen3.5-122B on Siona's two V100s: 122/136 on the Axonis Pty Limited judge papers first try (124 projected after a maths rerun), perfect IQ, maths and physics papers, about 49 tok/s, 61 °C, fully on GPU. Muse still leads at 125. @Alibaba_Qwen
Omen laptop tests (GTX 1660 Ti 6GB) on Axonis Pty Limited judge papers: gpt-oss-20b led reasoning (MATH-12 11/12, long doc 12/15) at 5.7 tok/s; Qwen3.6 13/13 tools at 9.6 tok/s, 76 °C; Gemma 4 E4B 45 tok/s with vision. huggingface.co/datasets/axon…
Earn 500 Microsoft Rewards Points Together with Me! edgemobileapp.microsoft.com/… #MicrosoftRewards via @MicrosoftEdge
My old GPUs got a big speed upgrade:
- 2016 Tesla P100s: Muse Glimmer 30B 5.2 -> 54.8 tok/s (~10x)
- 2x V100: Muse 52.4 -> 71.8
- Dell 2x GTX 1070: Qwen 3.8 27B 9.1 -> 25.3
Muse also scored 125/136 on my nine test papers, beating a 120B Nemotron (121) at about 4x the speed.
Results: huggingface.co/datasets/axon…
#LocalAI #LocalLLM #HomeLab
Update: Muse 2 still leads at 125/136. New challenger Siropia (Qwen3.8-Flash-Next on the open Strata engine) scored 121/136 at 86 tok/s on the same 2x V100s, about 7x faster than Nemotron. huggingface.co/datasets/axon…
Muse 2 still leads our nine judge papers: 125/136 at 53 tok/s on 2x 2017 Nvidia V100s at home. New challenger Siropia, our name for Qwen3.8-Flash-Next (IQ3_S quant, run on the Strata engine), scored 121/136 at 86 tok/s. Raw results:
huggingface.co/datasets/axon…
For context, Kai tested the same Qwen3.8-Flash-Next on the open Strata engine and got ~94 tok/s on a brand-new RTX 5070. Our 2017 V100s did 86. youtube.com/watch?v=mED-bodd…
New music video: We Make It Work 🎬 Every clip was generated and edited at home on our own hardware, with no cloud video services. Axonis Pty Limited
#AIMusic, #MusicVideo
Old datacentre cards, new tricks ⚡
Muse Glimmer 30B on two 2016-era 16 GB P100s went from 5 to 55 tokens a second, more than 10x faster, with prompt reading 2x faster too.
What did it: a P100-tuned llama.cpp build, both cards working on every layer together, and shorter draft guesses (2 instead of 16).
Same box, same model. Local AI from Axonis Pty Limited.
#LocalAI #LocalLLM #HomeLab
Credit where it's due: Muse runs on an open model from @AIatMeta. Open weights plus old datacentre cards and a bit of tuning gets you 55 tok/s at home. 🙌
Hey! I've been earning gift cards just by browsing with Edge Mobile. Download it with my link and we'll both get 500 Microsoft Rewards points!
edgemobileapp.microsoft.com/…
The Hollow Hour 🎃
A Halloween drum and bass track for tonight. Fog on the coast, the moon up, and something waiting in the trees.
Turn it up, lights off.
#Halloween #DnB #DrumAndBass #HalloweenMusic
Two old 8 GB GTX 1070s in a Dell R730xd, running Qwen 3.8 27B (4-bit) locally, just edged out Qwen 3.6 35B on dual RTX 3090s in our in-house tests at Axonis Pty Limited.
Nine test papers: 116/136 vs 114/136
Hardest set: 78% vs 67%
About 9 tokens a second, so it's not fast, but newer models are making old hardware useful again.
#LocalAI #LocalLLM #HomeLab.
