Pinned Tweet
AMDots is my preview for an AMD powered local agent, still in early development. Running entirely on AMD Ryzen AI Halo 395.
Before sharing the new AMDots demo, a few corrections to my October 23 video.
The generated explainer had errors I should have caught before posting:
The en AI Halo. “Gordon Halo” was incorrect.
The specifications mixed two configurations. The Max+ 395 system used in that demo has 128 GB of unified memory and an NPU rated up to 50 TOPS, rather than the spoken “5 TOPS.” The newer Max+ PRO 495 Halo has 192 GB and up to 55 NPU TOPS. AMD specifications.
The software name is ROCm. The narration mispronounced it and garbled several numbers, including “one hundred ninety-two gigabytes.”
The output fell short of the primary-source checking I requested and missed the requested animated diagrams.
The earlier voice stack used both CPU and NPU; the agent model ran on the Radeon GPU. The narration errors were problems in my software’s output and my review of it.
Since then, I’ve kept building.
The new demo introduces an upgraded LFM3 voice stack and CIRU’s optimized Qwen3.8-Flash-Next as the agent model:
-125B parameters+51B n-gram embeddings, with 6B active per token with a 4b MTP drafter
-256K native model context
-CIRU’s compact weights and custom runtime optimized for AMD Strix Halo
-Qwen architecture specifications · CIRU model and runtime
Better voice, more capable agents, and better research, from talking through a task to following sources and producing something useful.
This is what keeps me excited about the Ryzen AI Max+ 395: its unified memory and Radeon GPU give a compact desktop room for a substantial local AI workflow, with the large embedding table paged from local NVMe.
I’m excited to show you how far AMDots has come. New demo below. 👇
AMD provided the Ryzen AI Halo used for this project.
Before sharing the new AMDots demo, a few corrections to my October 23 video.
The generated explainer had errors I should have caught before posting:
The en AI Halo. “Gordon Halo” was incorrect.
The specifications mixed two configurations. The Max+ 395 system used in that demo has 128 GB of unified memory and an NPU rated up to 50 TOPS, rather than the spoken “5 TOPS.” The newer Max+ PRO 495 Halo has 192 GB and up to 55 NPU TOPS. AMD specifications.
The software name is ROCm. The narration mispronounced it and garbled several numbers, including “one hundred ninety-two gigabytes.”
The output fell short of the primary-source checking I requested and missed the requested animated diagrams.
The earlier voice stack used both CPU and NPU; the agent model ran on the Radeon GPU. The narration errors were problems in my software’s output and my review of it.
Since then, I’ve kept building.
The new demo introduces an upgraded LFM3 voice stack and CIRU’s optimized Qwen3.8-Flash-Next as the agent model:
-125B parameters+51B n-gram embeddings, with 6B active per token with a 4b MTP drafter
-256K native model context
-CIRU’s compact weights and custom runtime optimized for AMD Strix Halo
-Qwen architecture specifications · CIRU model and runtime
Better voice, more capable agents, and better research, from talking through a task to following sources and producing something useful.
This is what keeps me excited about the Ryzen AI Max+ 395: its unified memory and Radeon GPU give a compact desktop room for a substantial local AI workflow, with the large embedding table paged from local NVMe.
I’m excited to show you how far AMDots has come. New demo below. 👇
AMD provided the Ryzen AI Halo used for this project.
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 tch 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.
10
3
34
1,133
I think this means my 9800x3d is dead.
I will have to replace it with 9950x3d2.
Won't arrive until 14th :(
I will have to replace it with 9950x3d2.
Won't arrive until 14th :(
4
0
6
911
I'm having a horrible time with @mercury. I trusted them for my Apple M5 ultra preorders and I've been running around with support for days trying to figure out how to pay apple with my own money.
Their software is ridiculous, it says I have a 20k per day limit but a 10k per month limit. So what's my real limit? I can't get an answer from them.
I had a 51k limit when I ordered a couple of Ultras with it in October but apple doesn't charge until they ship, and mercury lowered my limit since then.
So now what I need to know, so I can manage my payments with apple and not lose my pre-order and get sent all the way back to March is:
If I convince Apple to let me charge 20k per day on my Mercury card, and pay it off every day, will it let me do this or will I be blocked by another limit?
When I pay it off every day how long will it take to give me my full credit limit back to use again, because if it's not fast apple will cancel my order.
If I can't get answers about how the bank works, I'll just need to move to a different bank where I don't have to worry about important payments going through.
Their software is ridiculous, it says I have a 20k per day limit but a 10k per month limit. So what's my real limit? I can't get an answer from them.
I had a 51k limit when I ordered a couple of Ultras with it in October but apple doesn't charge until they ship, and mercury lowered my limit since then.
So now what I need to know, so I can manage my payments with apple and not lose my pre-order and get sent all the way back to March is:
If I convince Apple to let me charge 20k per day on my Mercury card, and pay it off every day, will it let me do this or will I be blocked by another limit?
When I pay it off every day how long will it take to give me my full credit limit back to use again, because if it's not fast apple will cancel my order.
If I can't get answers about how the bank works, I'll just need to move to a different bank where I don't have to worry about important payments going through.
8
0
5
818
Interesting solution for agent native office tools. I want to see how I can use their SDK for some of my automated businesses workflows.
As far as I know, there is nothing else like this.
Anyone tried it out yet?
As far as I know, there is nothing else like this.
Anyone tried it out yet?
Today, we’re announcing Univer 1.4 and introducing Space.
With Univer SDK, build your own spaces with file management and document, spreadsheet, and slide editing.
Build an agent-native workspace. Build apps for your business.
Here’s a closer look. 🧵
0
0
1
338
Ciru.ai - Crown 👑 retweeted
Goodhart's law: once a score becomes the goal, it stops measuring the real thing.
3
4
63
2,745
Ciru.ai - Crown 👑 retweeted
In this clip from our conversation on the Strix Halo Stars podcast, @usmanpirzada reflects on the progress from Llama 70B to Qwen 35B and the developers making local AI more capable on @AIatAMD Ryzen AI Max+ 395. Listen to our full conversation: strix-halo-stars.com/episodes/1/
1
3
16
749
This is my custom engine that I'm working on with my M6 preparing for when I get my ultra hopefully this month.
These numbers are on lfm3 1.7b
@JoshuaSWarren
These numbers are on lfm3 1.7b
@JoshuaSWarren
1
0
4
639
Preparing for my m5 Ultra cluster that arrives later this month with an M6 Mac Mini. It only has 12 cores compared to the 80 on the Ultra. It should scale for prefill and the numbers could be very nice.
Interested to know what the spark is getting on LFM3 1.7B.
This is a $1000 ech2Wild @TechMDAI @liquidai
Interested to know what the spark is getting on LFM3 1.7B.
This is a $1000 ech2Wild @TechMDAI @liquidai
2
0
5
689
The 512 Mac Studio is going to be worth 40k+ each by April, mark my words.
They are going to good hands.
Congrats @ashxhart
They are going to good hands.
Congrats @ashxhart
So many people contributed to raise £24000 in less than a day.
Well done everyone!
I'm so glad we got to show Ash how much we depend on the work he does for free!
And soon we'll all reap the benefits of our investments 🙂
2
0
8
686
Lots of new small model releases, hard to know if they are worth the hype.
I tested some of the more promising looking ones on my FAST54 benchmark. It's an open source combination of eval questions that my models have struggled with the most, basically just take the hardest te and leave out the ones that all the models get right anyway.
Key takeaways:
Nanbeige 33B did the best on the tests most other even larger models failed.
Ornith 1.5 35B is a very solid contender across coding tasks.
The new SenseNova loop 9b actually does perform well above it's weight class. Need to do more testing here.
The tests are not perfect because the machines are always busy, so treat this as a ballpark not a scientific measurement. They have some different serving conditions, etc.
full tests here:
ciru.ai/research/fast5…
Download FAST54 benchmark:
github.com/ciru-ai/contes…
I tested some of the more promising looking ones on my FAST54 benchmark. It's an open source combination of eval questions that my models have struggled with the most, basically just take the hardest te and leave out the ones that all the models get right anyway.
Key takeaways:
Nanbeige 33B did the best on the tests most other even larger models failed.
Ornith 1.5 35B is a very solid contender across coding tasks.
The new SenseNova loop 9b actually does perform well above it's weight class. Need to do more testing here.
The tests are not perfect because the machines are always busy, so treat this as a ballpark not a scientific measurement. They have some different serving conditions, etc.
full tests here:
ciru.ai/research/fast5…
Download FAST54 benchmark:
github.com/ciru-ai/contes…
2
0
10
980
