Joined April 2026
- Tweets 99
- Following 249
- Followers 41
- Likes 861
5 Photos and videos
Michael Luo retweeted
Jun 4
If you are Taipei this evening, come to our AI agent meetup! Learn about what’s announced at Computex and their implication for open source!
luma.com/soczn0pg
PS. We have free Taiwanese fried chicken sponsored by @Docker
1
2
6
155
Jun 3
Come see us tomorrow!
Got 93 builders registered for tommorow's Taipei Agent Night meetup (June 4th, 6:00 - 9:00 PM). The energy is going to be unmatched.
We cover:
-Agents that act: OpenClaw, NemoClaw, and the robotics-meets-LLM scene.
-Agents that talk: Voice-first open stacks (EchoKit) you can interrupt and argue with.
-Agents that build: Coding agents and hardware like VibeKeys.
-The harness around them: Tools, memory, and self-evolution loops.
Expect live demos you can poke at (and probably break), plus hallway conversations you won't get anywhere else.
🎙️ The Speaker Lineup:
Kris Murphy @NVIDIAAI: Deep dive into Nvidia's Open Source AI (NemoClaw, OpenShell, Nemotron).
Michael Luo @luocompute @Olares_OS: Running your own AI datacenter on your desk with the RTX 5090M Olares One.
Michael Yuan @juntao @Docker Captain: Safeguarding agents and isolating tools with Wasm, containers, and NanoClaw.
Hydai @hydai_tw @secondstateinc: Interacting with AI agents via EchoKit.
HsinHo Yeh (Footprint AI): AI-native infra and Linux environments for agents using Containarium.
Alex Lau: Building MoJoAssistant and the scaffolding between models and the real world.
Grab one of the final spots here: luma.com/soczn0pg
6
69
Michael Luo retweeted
May 30
If you are in Taipei for ComputeX, register & come to our meetup. We will discuss how to run agents on your AI PC!
luma.com/soczn0pg
Featuring speakers from @nvidia @Olares_OS and local startups! Free pizza and Taiwanese fried chicken courtesy of @Docker 😂
2
6
5,019
Michael Luo retweeted
May 21
We are honored to share that Olares has won a 2026 BEYOND Award! 🏆
As one of Asia’s leading tech innovation platforms, BEYOND Expo @beyondtechexpo is the perfect place for us to demonstrate the next step in personal cloud computing.
If you are in Macao from May 28–30, join us to see firsthand how Olares One powers local AI agents.
📅 May 28–30 | Venetian Macao Cotai Expo See you in Macao! 🇲🇴
blog.olares.com/join-olares-…
#BEYONDExpo2026 #OlaresOne #LocalAI #BEYONDExpo #DIGITALPHYSICAL
2
7
192
Michael Luo retweeted
May 14
We are excited to share that NVIDIA NemoClaw now runs on Olares OS!
You can now run your entire AI agent stack locally and securely on devices ranging from the Olares One desktop to NVIDIA DGX Spark systems. @NVIDIA_AI_PC
Learn more:
blog.olares.com/olares-os-no…
2
6
23
3,736
May 13
We at @Olares_OS support your rights to repair
May 13
For anyone looking to modify Olares One hardware, DIY upgrades are fully supported.
This video breaks down the entire process of safely disassembling the device.
Discover exactly how to install an extra SSD, swap out the RAM, and reassemble the machine for maximum performance. 👇
youtube.com/watch?v=vFDm-mG9…
2
5
89
If you’re coming to Taipei for COMPUTEX 2026, we’re hosting a side meetup on Open Source AI Agents, near Taipei World Trade Center.
This will be a casual builder-focused gathering for people working on LocalAI, open-source models, AI agents, personal AI infra edge AI, etc We’re also opening a call for speakers. If you know a great project or speaker who will be in Taipei during COMPUTEX, please tag them or send them my way!
If you are building an interesting open-source AI project, working on agent frameworks, deploying local models, or exploring how AI systems can run closer to users and devices, I’d love to hear from you. Pizza and drinks will be provided.
1
1
5
127
Michael Luo retweeted
May 7
Max out your #NVIDIA #DGXSpark with Olares OS.
Don't waste hours on environment configs. #Olares turns your Spark into a dedicated personal AI cloud.
Enjoy fast AI model deployments, out-of-the-box secure routing, and effortless VRAM management.
Check out the docs for full details: 👇
docs.olares.com/1.12.5/manua…
3
8
396
Michael Luo retweeted
May 6
Check out this new video from LTH showing off a great use case for Olares One!
By pairing Cline and Continue in VS Code with local models running on Olares, he unlocked all the speed of AI code analysis and generation while keeping his codebase completely private.
Watch the full video:
youtube.com/watch?v=P89MNeIL…
1
5
285
Michael Luo retweeted
i don’t mean to be rude but some of your guys opinions about ai seem incoherent but more importantly , profoundly alienating to anyone who does not possess the incredibly narrow context and special language required to follow. which is kind of the core comms “problem” with ai
34
28
576
33,472
Michael Luo retweeted
May 1
This Workers' Day, work for yourself, not the clouds.
1
2
6
223
Michael Luo retweeted
Apr 27
Give your local AI a voice with Speaches.
This OpenAI-compatible server handles Speech-to-Text, Text-to-Speech, and real-time Audio Chat alongside Ollama. It comes pre-loaded with high-quality models like Whisper and Kokoro, ready to use immediately.
Speaches is #AvailableOnOlares. Install now.
market.olares.com/app/market…
1
7
193
Michael Luo retweeted
Apr 24
guys I am ready to vibe code
34
7
247
11,403
Michael Luo retweeted
Apr 24
We are excited to share that #Olares is joining the #NVIDIA Inception program!
We are thrilled to unlock NVIDIA’s cutting-edge developer resources to further enhance the stability and usability of Olares.
Let’s keep building!
#NVIDIAInception #AI
2
2
17
527
Apr 24
Excited to build more with @NVIDIAAI
Apr 24
We are excited to share that #Olares is joining the #NVIDIA Inception program!
We are thrilled to unlock NVIDIA’s cutting-edge developer resources to further enhance the stability and usability of Olares.
Let’s keep building!
#NVIDIAInception #AI
1
3
215
Michael Luo retweeted
Thor Olares MBP Pro Max all ready to install !
3
1
7
143
Michael Luo retweeted
You should buy an @Olares_OS
The team is super responsive. Any question (even stupid) I ask get answers in minutes.
The kids are doing what give us all in MBP maxi a way to have a CUDA box that can live in the backpack. Basically a super that we can also use for gaming !
1
1
5
362
Michael Luo retweeted
Apr 22
nyc built one of the most effective Space Pirate-flavored AI cabals on the planet, the most cracked secret society in America very quietly. what happens next will surprise you
5
1
60
4,051
Hardware-software co-design has been pushing LLM inference to toward specialized compute per stage:
Vera Rubin for Prefill, Groq LPU for decode. The theoretical performance gains are enormous.
But in practice, there's a major bottleneck:
Datacenters are built around homogenous clusters, so full specialization would require us to go cross-cluster.
With standard attention, The KV cache grows linearly, making network bandwidth a bottleneck very quickly.
However, hybrid models drastically change the equation due to one simple fact: Constant KV size.
Moonshot did the math:
At 128K context, standard attention could sustain 2 requests at a time, but hybrid easily did over 20. Thats an order of magnitude improvement.
Its the difference between heterogenous P/D disaggregation being a pipe-dream vs actually viable.
Apr 18
We push Prefill/Decode disaggregation beyond a single cluster: cross-datacenter heterogeneous hardware, unlocking the potential for significantly lower cost per token.
This was previously blocked by KV cache transfer overhead. The key enabler is our hybrid model (Kimi Linear), which reduces KV cache size and makes cross-DC PD practical.
Validated on a 20x scaled-up Kimi Linear model:
✅ 1.54× throughput
✅ 64% ↓ P90 TTFT
→ Directly translating into lower token cost.
More in Prefill-as-a-Service: arxiv.org/html/2604.15039v1
5
13
195
33,404