Want to help build a more regenerative and ethical AI future?
Australia
Joined January 2026
- Tweets3.1K
- Following198
- Followers318
- Likes1.8K
Pinned Tweet
Lots of new followers over the last week, for those of you who have joined recently - I'm Josh. From Australia.
I like to tinker with small language models on edge hardware - primarily in my field of passion agriculture 🌿
I'm an ambassador for @OpenBMB and look forward to fine tuning more models from their line up especially their vision models.
I'm currently building a custom front end and modified agent harness based of @NousResearch Hermes - this will provide an easy to use business personal assistant for farmers and food producers/manufacturers.
I'm now getting my head around the physical AI space with much enthusiasm.
I believe in ethical AI systems, I am a heart on my sleeve open source user and advocate across all devices I use.
If you want to be on the Right Side of AI - follow along and interact on my journey!
Do I start buying 12GB 3060s + cheaper DDR4 RAM when I can afford it, and build my VRAM + Memory empire there? Eventually scraping enough hardware to run Qwen3.8 27b?
Thoughts? Or Tesla cards? Who's doing local AI on the cheap and scrappy?
Still keen on this idea. X Algo link me with the Australian builders.
@PaulGug @NeoAIForecast put the Batman in the sky and let's make it happen.
Any other Aussies I've come across, I'm sorry I don't remember your handles 😅😅
🇦🇺🇦🇺🇦🇺🇦🇺🇦🇺
Australians who like AI and who like to build and tinker. We should plan a catch up somewhere?
My local council actually funded a vibe coding camping weekend recently - I unfortunately couldn't attend.
Or maybe we can start with an Australian based hackathon? If you are keen comment and tag other fellow Australians 🇦🇺
Josh retweeted
📱 35B parameters. Real-time inference. On a phone.
Meet Lookahead Expert Scheduling (LES), the inference system powering MiniCPM-V 4.7’s 35B-A3B MoE model on edge devices.
🚀 On an iPhone 18 Pro Max, the full system reaches 20.51 tokens per second, a 6.07× speedup over loading expert weights on demand.
How? Three coordinated optimizations:
- 🔮 Scheduling: Predict and prefetch the next layer’s experts while the current layer computes.
- 💾 Storage: Combine contiguous expert layouts with mixed-precision quantization to reduce loading overhead.
- ⚡ Compute: Use DSpark speculative decoding to verify multiple candidate tokens in one pass.
Predictions guide prefetching; the model’s actual router still determines which experts execute. The architecture, expert counts, and routing remain intact. Mixed-precision quantization reduces the weight file by 26.7%, with just a 0.16-point drop in the five-benchmark average.
LES shows how better inference systems can bring larger, more capable models to everyday devices. 🌍
👉 LES: github.com/openbmb/llama.cpp…
🔗 GitHub: github.com/OpenBMB/MiniCPM-V
#MiniCPM #OnDeviceAI #MoE #EdgeAI
Anyone else rocking the DAP? Back in my day they were just called a mp3 player. I really enjoy what FiiO has provided here.
It's a half way between an mp3 player and a phone but with no distractions. I only have two apps installed on it which is the music player itself & Metrolist for when I want to stream.
Distraction free high fidelity music listening.
has anyone ever experienced the absolute chaos that is alibaba cloud AI?
they have 4 websites for qwen, they have 3 different coding plans depending on where you live, they have straight limits and context windows, limited tool calls on a variety of uses. 84 different models spanning across different era's of qwen.
but they give away lots of free tokens... 90days from sign up and you get 1 million tokens for every model they offer - including image + video. it's where I started when i first dipped my toes in to AI. I was a qwen boy.
thanks @Alibaba_Qwen for the chaos and generosity.
pantler agent is in the github wild
next step - easy installer for windows and mac users who have zero technical knowledge
landing page
find customers and work out inference providing built into their subscription.
github.com/RegeneratusLabs/p…
Next @OpenBMB project - materia.
MiniCPM5-2B for reasoning and MiniCPM-OCR-1.0 for reading documents plus a 44M @OpenMed_AI PII encoder.
A privacy focused naturopath and integrative medicine practitioner and patient assistant.
This model and encoding stack will assist practitioners and patients to analyze their blood tests and symptoms from a more natural and integrative approach. Keeping all private patient data out of the cloud being designed to run on a consumer-grade laptop. Encoding ensures minimal model hallucination compared to using a cloud-based provider.
Australians who like AI and who like to build and tinker. We should plan a catch up somewhere?
My local council actually funded a vibe coding camping weekend recently - I unfortunately couldn't attend.
Or maybe we can start with an Australian based hackathon? If you are keen comment and tag other fellow Australians 🇦🇺
The wave of specialised models are incoming 🤓
This 4B model writes 3D scenes you can run in your browser.
CCORE’s Qwen3.5-4B fine-tune generates HTML and JavaScript using Three.js: temples, villages, interiors and animated scenes.
It runs locally, with a Q4 GGUF download of just 2.71GB.
Still an early preview, but the creator reports better scene quality and more successful first attempts in newer checkpoints.
Full FP16 weights, training data and an evaluation report are planned once training finishes.
Take me back to China. I want to set up somewhere nice with my laptop and build all day 💻
The vibe spots were on point and I imagine the deepseek latency would be low.
My hermes-band has now stepped into the world of rap...
Written, created and performed by Hermes.
Skill being managed by Hermes here:
github.com/RegeneratusLabs/h…
No cloud API for any music creation - all created on my machine. Having too much fun.
Behind the scenes video of the Hermes bot band funk record.
Maybe Hermes Bot Band could feature at the upcoming @NousResearch NousCon - land their first gig?
Repo:
github.com/RegeneratusLabs/h…
pantler agent coming along nicely.
almost ready to be released to the world - open source, sovereign based agentic software.
think personal business assistant geared toward farmers, food producers and food distributors/sellers.
connected to your existing services, this is an agent that understands your business, grows with you as you use it more. remembers with memory, learns new skills and capabilities and includes relevant capabilities relevant to the food industry. helps you with all aspects of business. built on my existing knowledge of the industry.
built on top of Nous Research Hermes - bringing simplicity to the complexity of AI agents - the joy of open source software.
let's get you out of ChatGPT & Claude chatbox and into AI that actually helps and grows with you.
Welcome to the Hermes bot band.
What's your thoughts on their first funk track?
Each Hermes bot is an individual band member - different instruments, personalities - all having to collaborate on one track together.
More details coming, in the meantime check out the hermes-band repo:
github.com/RegeneratusLabs/h…