Testing Qwen3.8-Flash-Next on my 64GB M4 Pro Mac mini via DwarfStar/Metal: 224K context, 284 tok/s prefill, 28.1 tok/s decode — nearly flat from 8K→224K. ~42GB model + SSD-backed N-grams. github.com/antirez/ds4/pull/…

well, this works better than Claude for Android or ChatGPT for Android has ever worked. It runs on my phone directly, no cloud vm. I can live edit it. Its multiplayer. And I can pick any provider/model. Just added artifacts. There is no moat. 24h of slop, developed on the phone, on top of Pi Durable :)

Today we launched Kolibri. On German National Day. A new LLM aleph-alpha.com/en/kolibri/ from Aleph Alpha available under Apache 2.0. Its been an intense few months across pre and post training to make this real. Look forward to getting adoption and feedback!

Dear @Aleph__Alpha team - thank you for making Kolibri-1 open. We care deeply about sovereign AI, and launching it on German Unity Day makes today feel especially fitting. As a small gesture of support, we’ve hosted and made Kolibri-1 free for anyone to try for the next few days. No GPU. No setup. Just try it. tesseracted.com/kolibri-1-ch… Open Models move us all forward. We’re rooting for you. 🇩🇪 Tesseracted Labs GmbH. tesseracted.com #alephalpha #openweightmodels

we are currently in the main frame days of this tech. i'm convinced it will reach the personal computer age, rather quickly. at which point we can own our shit again.

Your intelligence. Your model. Let SovAPI route it, or deploy a specialized model for your task.

Instead of cutting nightly software releases, I now have a nightly AI slop factory...

germany just dropped a sovereign open weight model kolibri by @Aleph__Alpha runs 3.5b of its 78b parameters per word and its math is kinda ridiculous: 96.9% on aime beats every mixture-of-experts model they tested, even 3x bigger ones. only a dense model doing 8x the work wins anyone can run it on their own servers, it thinks in german, and in their evals it tops every open model its size in english and german models read text in chunks called tokens, and i ran kolibri's chunker (its tokenizer) on the german constitution: it needed 15% fewer tokens than gpt-5's for the same text. "bundesverfassungsgericht" is 6 tokens for gpt-5 and 2 for kolibri. fewer tokens means cheaper, faster german, and more of it fits in what the model can read at once how it works, simply: 1. every layer has 384 tiny specialists, and a router sends each word to 6 of them. so it thinks like a 3.5b model, but it needs the memory of a 78b one: about 78 gb, which means 2 big nvidia gpus (h100s) or 1 h200 2. most layers only look at the last 512 tokens, and every 5th layer looks at everything. that's how it can read 1 million tokens (a few thick books) without it costing a fortune 3. it reasons in german. their team found that a little german reasoning data is worse than none: the model's german thoughts go in circles and never finish. so they made about 800k german reasoning examples and gave it a lot 4. it's trained to say "i don't know". they play a game with it where parts of the documents are hidden, sometimes to help it and sometimes to hide the evidence, and it has to tell which. when it didn't know an answer it admitted it 44% of the time. qwen3.5 did 11% where it's weaker: answering from memory, using tools over a long back and forth, and coding agents, where qwen models are ahead. and to run it you need aleph alpha's add-on for vllm, a popular open source server for running models if you have german documents and need to keep them on your own hardware, this is a big deal. huge congrats to everyone at aleph alpha, my good friend @MichaelLHofmann included!! i wrote up how it works, the benchmarks, how to run it and when to use it: tej.as/blog/aleph-alpha-koli…

Small bird, fast wings, Kolibri is here. 78B parameters. 3.46B active. Up to 1M tokens of context. Built in Europe. Now the weights are yours. Run it on your own hardware, under Apache 2.0.

Let’s explain the red tape playbook and why this is important. Notaries are well aware they add costs and inefficiencies yet want to keep the cash flowing Their best friend? Opacity. That’s where the opacity game plan kicks in: 1. Hide the “decision process” that leads to reintroduce notary regulatory capture by Member States at @EUCouncil a. Hide -against the law- all working documents of @EUCouncil b. Hide which “Member State” is siding with notaries regularory capture. c. Reframe monopoly cartel as “preventive law” 2. How to derail this. a. Repost and ask for @EUCouncil to stop breaking the law b. Ask for full transparency on who exactly is protecting the notary special interests: the names and titles of the people, the basis This was predictable : so when they say they talk “on behalf of Member State” they know they lie. The official studies across Europeans done by @EU_Commission @EU_Justice show founders and shareholders do NOT want outdated notarization. And the best way to end any debate : make it optional: if notaries or their special interests friends are so sure people want them, people will keep buying. They’ll vote with their money c. EU Courts have established that notaries are NOT public authorities. Most of their rights they often call “acquis” were created and reenforced during absolute monarchies and dictatorships to control the population’s activity. 3. Now they see EU Inc having momentum - 🇩🇪 notaries call it Autocracy Inc- their plan B is “we’ll let it formally exist but we’ll cripple it into meaningless and uselessness. ” For example Member States so far systematically refuse to monitor kpis by law and to allow citizens and institutions to monitor its success. Indeed if there were allowing this it would give the spotlight on their successful crippling of EU Inc, which to do a good crippling job needs opacity. A good example is the company law digitalization directive 2019/1151. Member States don’t ask nor want to communicate its effectivenes. Notary chambers refuse to answer. In Member States with notaries in company law, it’s probably less than 1%, maybe 5% of truly digital (remote, no in person meeting) notarization. And even within this much more expensive and slow than notary-free. They bet on Europeans moving to another topic. They bet on “we did create EU Inc, why you complain?” It’s our job as Europeans to make them lose their bet. Speak up. Amplify. Repost Let’s turn the lights on Fiat Lux.

Your model. Your GPU. Qwen3.8-27B on European GPUs, same price per token. Join us.

笑ってしまった

I don't understand the @MiaAI_lab drama. Are you upset that you thought she was a woman called Mia but she's actually a dude? I'm sorry but who gives a f*ck?

なるほど。VRAMにすべて収まる場合はllama.cppで、そうじゃない場合はStrataということか

This is one of the coolest local AI projects I’ve seen recently: Backburner lets your iPhone help your Mac run Qwen3.8-27B locally over a 10 Gb/s USB-C cable. The phone isn’t just storage, it actually participates in inference. github.com/StayLameBro/backb…

following so many local AI peepz here made me totally compute FOMOing... :D

To celebrate this unusual event, I'll be offering 25% off my book for a limited time. Use stefanmaier-25 in the coupon code. The book is literally how I grew my account! Get it here: mia-ai.net/book

Old CPU support comes in next week!

With the current progress in local AI efficency and tps speedups I can not keep up with the downlaods & calibrations... 😂😂

M5 Ultra + DwarfStar + pi + qwen3.8-flash-next q4 is working better overall for me. Effort high works well, I see less back and forth during coding and less errors in general. Now I'm optimizing performance, I'll then retry building something with it.

Next project: