earns thot a one-shot chat.

frg holds if daily gains show up on real work, not demos.

MLX-Serve 26.10.1 is out: speed across the board. Exactly faster. Qwen3.8 27B with its drafter +66% on an M5 Ultra, +28% on an M4 Max, +37% on an M1 Pro. Qwen Flash Next: +11% decode on M4 Max, +51% prefill on M5 Ultra. (up to ~5400 PP!!!!) Gemma 4, Qwen, LFM2.5, Spark, Bonsai 2 all faster, output byte-identical to 26.9.6 (across 18 models). New: * GGUF models run on our own MLX engine, written in Zig and Metal, instead of the llama.cpp fallback. (experimental) * 🍣 Sushi-format Flash Next packs run directly. * Qwen-Image 2.1 edits pictures from instructions. * Drafting is on by default for every model that supports it. Huge thanks to everyone who filed, tested and fixed this one: @STRML_, @sbusso, @wutang_superfan , lborloz, LXD-8 @alinselea zeeshanhaque21 @Lojza3D, h9q2cyxvgm-ui @kennethrdegraff, @CowboyCoderHQ , codysk, @Beamsters1, @CerebralCoding_, @AjAbsaki , and more.. THANKS ! Please comment if I did not include your name, for visibility, I know people mostly by their GH handle. Special thanks to @ashxhart and @Spangler3000 for pushing ! Benchmarked and tested across 5 machines: htmlhost.jax.workers.dev/ren… GH Release Download + Changelog: github.com/ddalcu/mlx-serve/… Website: mlxserve.com

Three ecosystems collide 🔥 AMD joins the MCDMA club, as we have now achieved true RDMA. @Apple @NVIDIAAI @AIatAMD We measured ~10 µs round trip bidirectionally. Huge thanks to @petruspennanen for giving me access to their Strix and Spark so I could get this up and running. The next test is running inference across them 👀 mcdma.dev github.com/ashhart/mcdma

ng oinflates open-source risk by skipping the messy parts.

If the model that runs your service stack lives… If the model that runs your service stack lives outside your jurisdiction, you don't have an AI strategy. You have a dependency with a nice UI. ↓ Full brief — carousel. #UAEAI

ship every agent in shadow mode first. full tool-call logs from day one. name one person who owns the failure queue before go-live. if you're building AI in the UAE, don't leave that queue to a shared inbox. #AIAgents #UAE #Dubai #LLM #GovTech

You're not building AI for the UAE if every token still leaves the country. I keep watching the same movie. Sharp agent demo. Happy stakeholder. Then legal asks where the files go and the project stalls for six weeks. Cloud APIs are fine for toys. They are a hard sell the second real client data, Arabic contracts, or internal SOPs enter the chat. Local and sovereign inference is what actually unlocks the work that dies in procurement. Cost stops being a roulette wheel — you pay for boxes and power, not a meter that explodes every time someone dumps a 40-page PDF into the prompt. Privacy stops being a slide. The data never leaves your desk, your free-zone rack, or the client's own floor. Control gets boring in the best way: swap the model, cut a bad tool, keep serving when the public endpoint has a bad afternoon. The gap I want more GCC builders to own is making local the default path, not the enterprise upgrade you bolt on after the third security review. Most teams still treat on-prem or private GPU like cosplay. The ones who put real inference under real workloads will win the trust fights everyone else keeps losing. If you are shipping agents or internal tools here and your stack still phones home on every request, you are leaving deals on the table for whoever won't. Who is actually running production loads on private metal or locked-down GPU in Dubai or Abu Dhabi right now? Drop the setup. I want receipts, not vibes. #UAEAI #LocalAI #SovereignAI

most govtech ai fails in the review queue, not the model. if the reviewer can't see the arabic source span beside the claim, they rubber-stamp. show evidence. design that ui before you ship the agent. #AI #UAE #GovTech #Dubai #Agents

113 tok/s think off. 70 tok/s at xhigh. 960 / 892 tok/s prefill at 64k and 258k. Mac Studio M3 Ultra 256GB, mlx-serve, nothing else loaded.

validate the payload before you call the llm. teams chase model drift for days when the real bug is a null crm field. agents amplify bad data, they don't fix it. if you're building agents in the uae, gate tool inputs first. #AI #Agents #Builders #UAE #Dubai

I just replaced a $180/mo Arabic LLM API bill with a model running on a desk in Dubai. Latency fell. Privacy stopped being a slide. Most "bilingual AI for the GCC" still ships every Arabic ticket to a foreign endpoint first. Names, invoice lines, sometimes partial ID numbers. Your DPO notices. Your customers don't love it either. I took mixed Arabic-English desk traffic we already had and pointed it at a quantized instruct model on hardware we owned. Same prompts, same few-shots, no cloud hop. Median turn went from roughly three seconds of round-trip to well under half a second on the local network. Cost per thousand replies is power and a bit of fan noise. What actually fixed the experience for real users here wasn't raw parameter count. Cloud models kept "correcting" Gulf Arabic into stiff MSA or flipping into English mid-answer when the ticket was clearly Arabic. I locked the system prompt to mirror the user's language and dialect and banned silent translation. Support agents stopped rewriting the bot's tone. Escalations on "sounds wrong" dropped hard in the first week we measured. If you run intake, ops, or customer desk work for UAE/GCC teams, give local inference one honest week on real tickets before you renew the SaaS. You get the speed. You get the data stay-home story. You get dialect control you can actually own. Anyone else running Arabic or bilingual models on-prem or on a UAE box? What broke first? #LocalAI #ArabicAI #DubaiBuilders #UAETech #SovereignAI

السيادة الرقمية ليست شعاراً سياسياً؛ هي شرط تشغيلي لنجاح… السيادة الرقمية ليست شعاراً سياسياً؛ هي شرط تشغيلي لنجاح الذكاء الاصطناعي في الإمارات والخليج. ↓ Full brief — carousel. #UAEAI

your agent doesn't need admin tool scopes. minimum perms only — same as a service account. keep write tools off until reads are stable. uae buyers will ask. ship least privilege day one. #AI #Agents #EnterpriseAI #UAE #Dubai

arabic retrieval dies on alef variants more than model choice. if you don't normalize أ إ آ ا and strip tashkeel before embed, your UAE rag is guessing. do that first. then chase the fancy model. #AI #UAE #ArabicNLP #RAG #Dubai

if you're building agents in the uae, version your tool schemas. one renamed field over the weekend and monday opens with silent fails. agents are api clients. treat the contract like one. #AI #UAE #Dubai #Agents #LLM

The next bottleneck for UAE AI isn't smarter models The next bottleneck for UAE AI isn't smarter models. It's who owns the runtime when those models sit inside real operations. ↓ Full brief — carousel. #UAEAI

tool results need schema checks too, not just the calls. an agent can report success and still drop nulls into required fields. one pydantic gate kills that class of bug. building for uae enterprise? validate on the way back in. #AI #Agents #UAE #Dubai #LLM

هذا موقع لشباب اماراتيين يتضمن وصفات القهوة وتقدرون تضيفون وصفاتكم عشان غيركم يستفيد وماتحتاجون تسجلون فيه حتى .. therecipedrop.ae/

Most AI systems pitched in the UAE would not survive their first real ops day. I keep relearning this every time a polished demo hits a live queue. The gap between AI hype and what ships in the UAE and wider GCC is not about smarter models. It's about everything the pitch deck leaves out. Clean English. Perfect sample data. No residency wall. No manager who still needs a paper trail. You clap in the room. Then Monday arrives and the system meets a bilingual finance pack and a tool chain that was never stress-tested. Here's what I learned running this in production. I used to grade AI by how sharp the answers looked. Wrong meter. The thing that changed my filter was tracking what dies before the model even gets a clean prompt. Intake and routing killed more pilots for me than weak generation ever did. Arabic-English switches mid-document. Dates in three formats. Files that looked structured until you opened them. Local constraints that block the cloud shortcut people assume is fine. So I flipped the order. Design the failure path and the local rails first. Human override. Logging you can audit. Intake that doesn't assume English-only perfection. Only then drop the model on top. That sounds boring. It's also why some teams here ship and most stay stuck in pilot theatre. Hype sells intelligence. Shipping in this market sells survival under messy, bilingual, sovereign-aware conditions. If you've put AI into a real UAE or GCC workflow, what broke first for you? #UAEAI #AIBuilders