Just release the 70B bitnet 1.58 model. Why don't you do research on that path instead of all this meh stuff?

⚡ 𝗠𝗶𝗻𝗱𝗯𝗲𝗮𝗺'𝘀 𝗟𝗶𝘁𝗲𝘀𝗽𝗮𝗿𝗸 𝗥𝘂𝗻𝘀 𝗧𝗲𝗿𝗻𝗮𝗿𝘆 𝗟𝗟𝗠𝘀 𝗼𝗻 𝗖𝗣𝗨𝘀 𝗮𝘁 𝟰𝟬 𝘁𝗼𝗸/𝘀, 𝗨𝗽 𝘁𝗼 𝟵𝟳𝘅 𝗙𝗮𝘀𝘁𝗲𝗿 𝗧𝗵𝗮𝗻 𝗣𝘆𝗧𝗼𝗿𝗰𝗵 Mindbeam open-sourced a Python library that runs stripped-down AI models on CPUs. Ternary models compress weights to just three values: negative one, zero, or positive one. That kills multiplication entirely, the software only does addition and subtraction. On an Apple M5 Max, a 2-billion-parameter model hits 40 tokens per second, which is 18 times faster than PyTorch. Memory drops from 5 gigabytes to under one. Energy use falls 6 times on Apple silicon and nearly 13 times on AMD. Output quality matches full-precision models. It currently supports BitNet and Falcon Edge models and installs with a single pip command. The code is on GitHub. #Mindbeam #Litespark #LLM #CPU #Inference #OpenSource ─── 🤖 𝗙𝗼𝗿 𝗺𝗼𝗿𝗲 𝗔𝗜 𝗻𝗲𝘄𝘀 𝗮𝗻𝗱 𝘀𝘁𝗼𝗿𝘆 𝘀𝗼𝘂𝗿𝗰𝗲𝘀, 𝘀𝗲𝗮𝗿𝗰𝗵 "𝗚𝗲𝗻𝗔𝗜𝗦𝗽𝗼𝘁" 𝗼𝗻 𝗧𝗲𝗹𝗲𝗴𝗿𝗮𝗺

🫱 之前推荐过 Bonsai 1bit 系列模型，这次再推荐 Bonsai Image 模型。 huggingface.co/spaces/prism-… 太酷了，每一款都是真正面向消费者的小模型。 🙄️顺便留一个清单，都是本地能跑的小模型： Bonsai-8B MiniCPM5-1B LFM2.5-1.2B-Thinking Qwen3-1.7B BitNet b1.58-2B-4T Gemma 3 1B

Gate 1 (survival): post-hoc ternary 3.38 bits/byte = dead. QAT = 0.21 = alive. 94% of the damage recovered. Gate 2 (scaling): the gap collapses 0.193 -> 0.029 as params grow x11. BitNet b1.58's scaling law, reproduced on RWKV's DPLR recurrence.

I'm totally onboard for this too. I wanted to crowd-fund an asic of the ternary transformer model created by Microsoft, BitNet but I've just been too busy to pursue it.

You are levels of retard beyond. You do know all the "people " feeding you this shit are bots right? That you jyst kught be one of thhe retarded human lynchpins that holds reforms imported ecuadorian bitnet tigether.. You fucking retard.

私は理科大時代にBITNETを使っていませんでしたので、.BITNETなアドレスはありませんでした。 私が卒業後の、留年した理科大の知人（BITNETユーザー）とのやり取りが、スライドのものでした。

An 8GB Pi 5 can run: • TinyLlama 1.1B • Gemma 3 1B • Gemma 3 4B • Qwen 2.5 1.5B • Qwen 2.5 3B • Phi-3 Mini 3.8B • Llama 3.2 3B • SmolLM2 • BitNet b1.58 2B Sounds like a great learning project to me

Bitnet weights are the only open source asset with real revenue. every other "open AI" is a vcfundraising pitch with a github repo

100% recommend Pewdiepie's AI harness: github.com/pewdiepie-archdae… Odysseus'll help you find the right FOSS self-hosted AI for your specific hardware. If you have a bad GPU but, a decent CPU I recommend Bitnet as it supposedly works best on CPUs. Frontier models are great for bleeding edge progress but, for anyone concerned about life, liberty, and happiness FOSS is the only way to go.

Based on my experience with Grok, GPT, Gemini, Qwen, Minimax, Claude, Deepseek... For something I'd keep on me, I'd want a locally hosted, probably whichever FOSS one I like at the time. At the moment? Probably Minimax, Kimi, or Qwen. Though now that I think about it Bitnet might be the better choice for the hardware constraints.

お楽しみ企画2：「京大がネットサービスのHubだったころ」～UUCP, NetNews, ftp, archie, BITNET… #関西インターネット老人会

Well, I was doing well with Claude Fable for a few hours of work then got bumped down to Opus while it was doing a security audit. The guardrails failing in this way triggers my standard operating procedure to seek latest alternatives. This is the open source repo for HumanityOS that Fable failed with github.com/Shaostoul/Humanit… I'll be checking out Odysseus by Pewdiepie for the harness github.com/pewdiepie-archdae… Then probably Qwen (like the old version I presently have) or whatever other free open source model that is out with good marks on complex dev. Hopefully there are good reviews on Bitnet now that it's been out for a bit.

ローカルで動かしているBitNetに「XXとは何であるか」という質問を投げたらだいたい妄想ハルシネーションを語るんですが、質問文に「XXはYYによってZZされ、ZZはAAのBBである云々」という前提を置いて「ZZは結局どうなったか」と訊ねると、けっこうそれっぽい答えを返します。

今思ったけど、クローズドLLMの一部が既にbitnet b1.58を既に採用していたりtransformer捨ててたりしたらおもろいなあ

🐍 BitNet ⭐ 39,291 stars **"BitNet.cpp : L'inférence ultra-rapide des modèles 1-bit enfin accessible !"** #GitHub

正解：3 解説：1ビットLLM（代表例としてBitNet b1.58など）は、モデルの重みを従来の浮動小数点ではなく、「-1、0、1」の3値（1.58ビット）で表現する最新のアーキテクチャです。