David King

@Youssofal_ That's incredible! Love your work on MTPLX. On an M3 Max 128GB and it's actually made Qwen 3.6 27b bearable.

All this made with a local model running on a laptop at 50 TPS with MTPLX

@korale77 @ivanfioravanti Interesting 👀

@ivanfioravanti Yes! Switch to high power, disconnect external monitor, reboot.

M5 Max in Automatic instead of High Power is much slower, I see at least a 15% difference. So when you really want to push... High Power mode on!

@mastra Saw this this other day and was super excited!! Thank you! We run lots of NestJS based API's at my company.

👀 ICYMI Drop-in Mastra module for NestJS.

@Youssofal_ @KrakowiakK And for me, I love Pi! Will test it out now, this looks amazing.

@KrakowiakK this one’s for you 😂

Oh and I forgot something else. Pi is now natively available in the CLI for coding work. Run: MTPLX Start and select PI It will connect or start a new Pi session.

@LyalinDotCom Followed because of @ivanfioravanti 😂

I need 1,400 more followers. Anyone here into AI topics? I post on this a lot. Thanks.

@ivanfioravanti Ivan just want to say thank you! Every morning I wake up to check your threads, keeping us on the bleeding edge. I own an engineering company and thanks to you I have transitioned ALL my developers to local AI with Pi.dev. Forever grateful.

3/3 M5 Max Details mtplx-qwen36-27b-optimized-speed MTPLX API Benchmark Results Hardware: Apple M5 Max, 128.0GB RAM, 18 CPU cores, 40 GPU cores 0.5k pp 815 tg 52 t/s 18.4GB 1k pp 699 tg 48 t/s 18.7GB 2k pp 639 tg 49 t/s 20.3GB 4k pp 650 tg 41 t/s 20.8GB 8k pp 604 tg 45 t/s 21.2GB 16k pp 512 tg 39 t/s 21.2GB 32k pp 370 tg 38 t/s 27.5GB 64k pp 260 tg 25 t/s 38.2GB 128k pp 187 tg 22 t/s 61.0GB

1/3 A quick update on MTPLS after testing release 0.1.6. I'll create a new article, with future updates fixing latest issues: Pro: - Memory issue fixed! - Text Generation top! Cons: - prefill gets super slow after 16K contexts. As you can see in table below. 128K with 187 t/s is eternity 😢 (168 on M3 Ultra 😭) - batch / parallel calls not supported, problem for coding agents

@jpgstoreNFT 🫡🙏

After serving the Cardano ecosystem since 2021 and sharing an incredible journey with thousands along the way, we’ve reached a point where the JPG Store & Comet platforms have to be sunset. While we deeply value the people who have supported us, the platforms have reached a stage where they are no longer sustainable to operate. JPG Store & Comet will continue to be online until the 23rd of May. After which, the website will be shut down, though the smart contracts will continue to live on. 🧵1/6

recommended viewing. one more time, on it's own. this is probably yhe most practical talk on using coding agents i've watched to date. watch it. by @lucasmeijer it's also a great demo of pi and captures exactly why i built it. youtu.be/fdbXNWkpPMY?si…

We’re excited to announce today that @mastra has raised a $22M Series A led by Spark Capital. This brings our total capital raised to $35M:

@calcsam @mastra AYE!!!!!!!

@samwhoo Exceptional in very way! 👏🏾 Masterfully written and fun to play with the interactives. Thank you!

I think this is the best post I've ever made.

@stevibe 👀... interesting! Looking forward to the local model results.

Everyone says GPT 5.4 and Claude Opus 4.6 are neck and neck. I found an angle where they're not even close. I built a spatial reasoning puzzle: 6 questions across 3 difficulty levels (2×2, 3×3, 4×4). Results: 🟢 GPT 5.4 — 6/6 (perfect) 🟡 Gemini 3.1 Pro — 4/6 (missed 2 hard ones) 🔴 Claude Opus 4.6 — 1/6 (3 hit the 10k token limit without answering, 2 wrong) But here's what's wild: GPT and Gemini solved each puzzle in ~2k tokens. Opus burned through 10k+ tokens on three questions and still couldn't produce an answer. Local models up next!

@xpasky My favorite was "Like me, he thinks open-source and open protocols are a necessity, not just lipstick on a corporate pig." 😂😂 Masterfully written. I laughed so hard.

Mario just continues to prove his impeccable taste and prioritizing that over any instant gratification. (lol "Despite its Tolkien-inspired name, Earendil is not a tech company with fascist tendencies")

@badlogicgames Congrats!! Excellent read, loved the background context. So many great things here but my favorite is: "open-source and open protocols are a necessity, not just lipstick on a corporate pig." FACTS!!!😂🙌 Long live the people of Earendil

People of pi. BIG NEWS. I've sold out. Let me know how you feel about this in the comments below. mariozechner.at/posts/2026-04-…

@Claynaldo_7 @FahadoukenNFT @cryptodramz @1080Lt @lennaverse @ClayJoons @lozzaclay @claybrobot @LallyMcClay ROFL!!! I'll send you one of my @officialmocossi blue chips 🤣

@theuttermost @FahadoukenNFT @cryptodramz @1080Lt @lennaverse @ClayJoons @lozzaclay @claybrobot @LallyMcClay Haha send me your blue chips 🤣

Thinking of names from the glorious NFT bullrun of 2021: @FahadoukenNFT @cryptodramz Cardano Jer CJ Theuttermost (David) Soupnsando Ivan @1080Lt Cardanosalad Jak Quantumpanda Luney @lennaverse @ClayJoons @lozzaclay @claybrobot @LallyMcClay Lucid Who did I miss?

@itsolelehmann 100% worth it. I have M3 Max 128GB and run MLX models. It's super fast with the 128GB I can even run very nicely Qwen 3.5 122b. I code all the time traveling completely locally. Only issue is battery so, have a second battery pack haha

is it worth it to get the m5max macbook pro? I want to be able to run local models. how much ram should I get? can't get a mac studio because I'm traveling a lot this year.

@Claynaldo_7 @FahadoukenNFT @cryptodramz @1080Lt @lennaverse @ClayJoons @lozzaclay @claybrobot @LallyMcClay I am, I am. You're my very first follower 🥹, maybe I should send you one of my NFTs from 2021 bull run haha

@theuttermost @FahadoukenNFT @cryptodramz @1080Lt @lennaverse @ClayJoons @lozzaclay @claybrobot @LallyMcClay All good bro, tell me about it. What a time to be alive. Hope you are doing well 👊🏻

@jaykishanD @badlogicgames 😂

@badlogicgames Mario, you forogot to add “Thank you for your attention to this matter” in your tweet.

big news tomorrow. BIG NEWS. the biggest of news.

@ivanfioravanti 😂

When you vibecode something, the important thing is not looking too much at code generated. Add tests, add docs, add diagrams, ensure it works, but don't look at that code!

@Claynaldo_7 @FahadoukenNFT @cryptodramz @1080Lt @lennaverse @ClayJoons @lozzaclay @claybrobot @LallyMcClay What up brother. I am honored to have made such a prestigious list! Those were some of the best days of my life!

@theuttermost @FahadoukenNFT @cryptodramz @1080Lt @lennaverse @ClayJoons @lozzaclay @claybrobot @LallyMcClay 👀👀👀