Clément Pillette

Multi-agents AI that listen, reason, and act — built for enterprise workflows with reliability and security at the core.

[Clément Pillette @ClementPillette]

I've been comparing two setups running Qwen3.5-397B-A17B at full 262K context:

🖥 Mac Studio M3 Ultra (512GB) — €14,500
⚙️ Custom workstation, 4× RTX PRO 6000 (384GB VRAM) — €45,000

Results:
• Workstation: 46.9 tok/s, 1,100W, 51 dBA
• Mac Studio: 35 tok/s, 120W, ~15 dBA

The Mac is 6.7× more energy-efficient per token. Over 3 years, the TCO gap is nearly €40K.

I have never been a Mac guy, but I have to admit that the Mac Studio is currently the most attractive hardware for running local AI agents.

[Clément Pillette @ClementPillette]

@LottoLabs What a couple of days for GPT5.6, that’s it 😅

[Clément Pillette @ClementPillette]

@The_Only_Signal @LottoLabs I have started the project of converting my air cooled workstation to water cooling. Will have a total of 6 RTX6000 Pro MaxQ in waterblocks, the Theadripper as well. I will have 2 MO RA IV 600 radiators that will be some meters away from my workstation to exhaust the air outside

[Clément Pillette @ClementPillette]

@simpsoka Would be I think a big plus if /goal could you a smart model router to decide the level of the model (M, H or xH) to optimize the consumption. I currently use it to run cycles of different scenarios, collecting traces, telemetry, QA data, so Codex improves the code between cycles

[Clément Pillette @ClementPillette]

@Dimillian 💯

[Clément Pillette @ClementPillette]

@ivanfioravanti Just bought 2 extra RTX6000 Pro, so I am really counting on the SM120 support

[Clément Pillette @ClementPillette]

@levie Are there some enterprises considering inference on premises? Some OS models can perform a good portion of agentic tasks. Then with a clever routing, only the very advanced tasks could be channeled to frontier models of the key labs.

[Clément Pillette @ClementPillette]

@embirico @Dimillian Since May 2025, I was using both Codex and Claude Code. As of November, I was using less and less Claude Code. In May 2026, I have cancelled all my Anthropic subscription as I wasn’t using Claude Code since March. I am sometimes using Droid to run my local models.

[Clément Pillette @ClementPillette]

@0xSero I am going to buy waterblocks to liquid cool my GPU , this seems to be an interesting approach, as you can more easily channel the heat out, or to another room in the winter, better thermals and less dust issues. I am also going to add 3 more RTX6000 Pro max Q

[Clément Pillette @ClementPillette]

@antirez @AMD Ubuntu 24.04

[Clément Pillette @ClementPillette]

@Dimillian I am hitting the weekly limits often within 5 days, so if Thibault doesn’t perfect frequently enough resets, I have have my weekend free 😂

[Clément Pillette @ClementPillette]

@ajambrosino Increasing my usage, though the weekly limit is allowing me weekends time to time

[Clément Pillette @ClementPillette]

@TeksEdge I can confirm it is an excellent TTS model, by far superior to other OS TTS I have tried.

[Clément Pillette @ClementPillette]

@MaziyarPanahi 🤔

[Clément Pillette @ClementPillette]

@MaziyarPanahi 😅. Indeed!

[Clément Pillette @ClementPillette]

@MaziyarPanahi Because they don’t trust CharGPT for data protection ?

[Clément Pillette @ClementPillette]

@ivanfioravanti @antirez Not all the benchmarks are benchmarkmaximable

[Clément Pillette @ClementPillette]

@ajambrosino Looking forward to have it running on Linux

[Clément Pillette @ClementPillette]

@Galanthai @ivanfioravanti Many corporations I know, are only providing Copilot. Employees aware of the power of the much more advanced models/ agents, are using theses secretly so they can gain productivity. Would be wiser if those corporations were adopting enterprise solutions from OpenAI or Anthropic

[Clément Pillette @ClementPillette]

@ivanfioravanti Did you try the Codex app? It’s also fantastic that you can run it remotely on your iPhone; I don’t feel obliged to stay the whole time behind my computer. Also, if for instance I am walking the dog and an idea pops up, I can dictate immediately so I don’t lose the idea & logics