Said | CeliaLabs

Custom AI automations and fine-tuning — for teams and solo operators. Runs on your hardware. Free recipes + tests posted here. DM "BUILD"

[Said | CeliaLabs @Celia_Labs]

You're probably losing hours every week to work a machine could do for you.
I build those machines — apps & automations that handle the repetitive stuff:
• Businesses: agents that qualify leads, answer support, kill data entry
• You personally: small tools that quietly run your day in the background

My edge: I can build them to run privately on your own hardware — your data never leaves the building.
That's CeliaLabs.
Got a repetitive task — at work or at home? Drop it below and I'll tell you if it's automatable 👇

[Said | CeliaLabs @Celia_Labs]

The wheels on the bus go round and round 🚌🎶
Qwen3.8-Flash --> GLM-5.3 Flash

[Said | CeliaLabs @Celia_Labs]

I love #OpenSource 😍
Thank you, China. 🇨🇳❤️
谢谢中国 🇨🇳

@MiniMax_AI @deepseek_ai @Zai_org @Alibaba_Qwen @Kimi_Moonshot

[Said | CeliaLabs @Celia_Labs]

Chinese Models on Chinese AI chips 🇨🇳 💪
congrats @Zai_org

[Said | CeliaLabs @Celia_Labs]

The honest verdict:

• Building cloud apps or handling high-traffic APIs? NVIDIA & CUDA are still completely unmatched.

• Local research, 100% data privacy, and testing giant 70B–405B models at home? The M5 Ultra + MLX setup is easily the most power-efficient, quiet powerhouse you can get.

Are you sticking with your NVIDIA GPU rigs, or is the new Mac Studio tempting you to switch? Let’s hear what you’re building! 💬

[Said | CeliaLabs @Celia_Labs]

The new Mac Studio M5 Ultra just quietly solved the biggest headache in local AI: running massive 405B models without a giant, power-hungry server rack.

With up to 512GB of Unified Memory and 1.2 TB/s bandwidth, it fits giant models into a single, silent box on your desk.

But can Apple’s MLX actually take on NVIDIA’s CUDA? The reality of hardware vs. software broke down in the replies 👇

[Said | CeliaLabs @Celia_Labs]

Hardware is brute force, but software is the real superpower:
his is where CUDA still rules:

• 15+ years of kernel fusion (FlashAttention-3, CUTLASS, Triton).

• Serving infrastructure like vLLM and SGLang (PagedAttention, chunked prefill, continuous batching).

• Cutting-edge speculative decoding methods (DFlash block-diffusion, DSpark dynamic verification scheduling) are designed around CUDA graphs and async GPU streams.

MLX handles standard Multi-Token Prediction (MTP) and speculative drafting cleanly, but its ecosystem is built for single-tenant research, not high-concurrency production serving.

[Said | CeliaLabs @Celia_Labs]

Qwen 3.8 Flash ...

[Said | CeliaLabs @Celia_Labs]

Just pulled the trigger on the new Mac Studio for local AI.
Can't wait to push Apple silicon and MLX to the limit.

#MLX #COREAI #Apple #LocalAI

[Said | CeliaLabs @Celia_Labs]

Every local AI lover right now…

The story isn’t over yet, check the replies for the rest 👇

[Said | CeliaLabs @Celia_Labs]

This is hilarious 😂😂
Finally someone explained the current AI Ecosystem

[Said | CeliaLabs @Celia_Labs]

@zebassembly Using everyday
Perfect for everyday needs
But Be careful it hallucinate a lot

[Said | CeliaLabs @Celia_Labs]

Everyone is arguing about which model to use.
Meanwhile the teams shipping working agents are spending their time on:
— tool interfaces the model can actually use
— what goes in the context window, and what doesn’t
— error handling and retries
— orchestration and task decomposition
The model is maybe 20% of the outcome. The harness is the rest.
(Until your task horizon gets long. Then the model becomes the ceiling again.)

[Said | CeliaLabs @Celia_Labs]

Open models closed 80% of the gap. I still pay for both.

I deploy local AI for myself and for clients. Privacy work, volume work, anything repetitive — that all runs local now, at zero marginal cost.

But I keep Claude Pro. Not for better answers — for fewer wasted passes on large codebases. Local models make me babysit context, and that babysitting is the real cost.

Gemini Pro I’d cancel. That’s a convenience tax on ecosystem lock-in, not a capability edge.
Buying back hours beats saving $20.

What’s still on your bill — and what’s keeping it there?

[DeepSeek @deepseek_ai]

DeepSeek-V4-Flash-Vision-Exp is now live on the DeepSeek API Platform! 🚀

🔹 This experimental multimodal model matches DeepSeek-V4-Flash on text capabilities—including agents, reasoning, and world knowledge.
🔹 On multimodal agent benchmarks, V4-Flash-Vision-Exp makes a major leap over V4-Flash, bringing multimodal agent performance close to Opus-4.8.

Try it with model='deepseek-v4-flash-vision-exp'. DeepSeek Harness 0.1.1 was released today with out-of-the-box support for the new model.

1/n

[Said | CeliaLabs @Celia_Labs]

@markproduct LC

[Said | CeliaLabs @Celia_Labs]

@aditikbuilds Lacal ai systems implementation

[Said | CeliaLabs @Celia_Labs]

DFlash 2 is the real deal.
Tested it on an M5 Pro MacBook Pro for coding workflows in OpenCode—easily outperforming both MTP and DSpark.

Game-changing acceleration for local models.
Thank you @zhijianliu_! 🔥

[Zhijian Liu @zhijianliu_]

DFlash 2 is here! Qwen3.8-27B at 70 tok/s on an M5 Max MacBook Pro.

⚡ Up to 4.6× the speed of autoregressive decoding, with the same output.

This is the next generation of DFlash, seeded at Z Lab and upgraded at Inco AI. Get one more accepted token on every pass, for free!

https://t.co/We0lwYPSBl

[Said | CeliaLabs @Celia_Labs]

Hey @Alibaba_Qwen, the chunky capybaras are cozy on the racks, but our edge devices are starving! 🦫⚡️

When can we get the official smol squad (0.8B, 2B, 4B, 9B)? Let the little guys run free on our laptops! 🚀💻 #Qwen #LocalAI

[Said | CeliaLabs @Celia_Labs]

I love #OpenSource 😍
Thank you, China. 🇨🇳❤️
谢谢中国 🇨🇳

@MiniMax_AI @deepseek_ai @Zai_org @Alibaba_Qwen @Kimi_Moonshot