We can have our own entirely local AI assistants for research and work. This is CIRU AMDots: a voice interface for a general-purpose agent team, with the AI running locally on one AMD Ryzen AI Halo. AMDots is your own little AI crew, brought to life as colorful dots. Give the main one an idea and watch its workers spring into action, exploring, researching and creating together. Keep chatting while they work, then see what they’ve made, from useful discoveries to a narrated video. The brain is MiniCPM 4.7 Ultra , CIRU HALO AGENT, my custom Q8 conversion and optimized serving stack built on OpenBMB’s vision-language MoE: 35 billion total parameters / 3 billion active per token 262,144-token model context limit Custom Ciru vLLM + AMD ROCm+DSpark On the 395: ~1,969 prompt tokens/sec and 40.5 generated tokens/sec with a roughly 64K-token prompt 122.8 aggregate generated tokens/sec across eight concurrent vision/tool requests Those are separate model-serving benchmarks; the eight-request figure is combined throughput across requests. Model, runtime and benchmark details. The voice stack uses the rest of the hardware too: GPU: conversation and worker-model inference CPU: Moonshine Small Streaming English speech recognition, 123M parameters CPU + NPU: Pocket TTS, 100M parameters, with recurrent feed-forward operations offloaded to the XDNA 2 NPU Speech synthesis stays off the GPU, leaving that compute available for the agents. Voice and workers share the language model, with live conversation taking priority. Local model inference, a visual agent workspace, and research that can become something you can watch and hear. All on a small AMD desktop. AMD provided the Ryzen AI Halo system used for this project.

Local voice cloning is getting really good. I tested the newest Higgs model against @AntLingAGI new Ming 3.0 voice stack. Higgs has been my go to until now, really great quality and performance on local hardware. After this test, I am going to switch to Ming 3.0. Even the 500m model sounds better than Higgs in most of my tests. I'm getting a new mic tomorrow, sorry about the sound quality. I'll make some more videos later where I try out the entire stack, including the new Ming-3.0-audio-flash, a 126 billion parameter long-form audio reasoning agent.

What we all crave. Kimi K3.1 Padoga. Reminds me of Astra Ultra padoda. Not quite as epic though. This took 9% of my weekly $100 plan limit.

While other banks are banning agents Mercury Bank is giving my agents their own credit cards. 😎

A one shot of padoga's in padogas. A bit of chaos from 3.8 flash running locally on my AMD 395. ciru.ai/padoga-factory/garde…

Another from scratch model today! This is much more significant than another remake of Qwen. Plus, this one from an American lab @Recursive_SI Speed out of the box is promising, no optimization done yet except for tuning the config. Quality tests come next. For reference, this feels faster than initial Qwen3.6 35b when it first launched, which we later doubled in speed after optimizing for AMD hardware. If the quality is there, this is a big deal for local inference!

Today is a good day. A South Korean company (behind gaming hits like PUBG and Subnautica) has released a completely new, original model architecture and a model trained from scratch, building on existing innovations and leading-edge methods. It is incredibly expensive and challenging to train a competitive foundation model from scratch. Even giants like Google, with vastly greater resources, are being beaten on important benchmarks by this gaming company's first release. This is a much bigger deal than people realize. It looks competitive and perhaps even superior to Qwen's similarly sized MoE in several areas, even giving the much larger active-parameter dense Qwen3.8-27B some competition. They released it right away with NVFP4 and Apple Silicon MLX versions, under Apache 2.0. I think we're about to see some massive competition in open models, and I'm really excited about this. I'm going to begin making optimizations and builds for this model immediately.

It's a joke. I'm pissed. Give me $300 actual credit off my bill instead of $1000 of those fake usage credits any day.

Loktar is right. Also, should go further. I always tell clients it doesn't matter if a model has some secret back door or whatever to try to steal your data or disrupt your operation. With sensitive information our models should not be given the ability to do that in the first place. Even an innocent model is capable of making mistakes and deleting sensitive information or releasing it to the entire world. When dealing with sensitive information you should have safeguards in place that don't trust the model in the first place to be sure that your information is safe. An innocent model accidentally leaking information or some sophisticated secret token path that phones home with data, neither of them should be able to successfully extract anything. You own the harness and you have complete control over your network and the software that runs on it.

I love this. Memory management is one of the largest problems with agentic fleets. Tencent released a model trained specifically for memory management, with their own stack. I definitely want to try this out. I've been looking for memory solutions for a while, it's not easy. @bradmillscan

I can't stand this and it makes me so happy to be using models locally.