Mastodon

st1nger :unverified: 🏴‍☠️ :linux: :freebsd:<a href="https://infosec.exchange/tags/Microsoft" class="mention hashtag" rel="nofollow noopener" target="_blank">#Microsoft</a> has introduced <a href="https://infosec.exchange/tags/BitNet" class="mention hashtag" rel="nofollow noopener" target="_blank">#BitNet</a> b1.58 2B4T, the largest-scale 1-bit <a href="https://infosec.exchange/tags/AI" class="mention hashtag" rel="nofollow noopener" target="_blank">#AI</a> model to date with 2 billion parameters and the ability to run efficiently on <a href="https://infosec.exchange/tags/CPU" class="mention hashtag" rel="nofollow noopener" target="_blank">#CPU</a> - It's openly available under an MIT license <a href="https://huggingface.co/microsoft/bitnet-b1.58-2B-4T" rel="nofollow noopener" translate="no" target="_blank">https://huggingface.co/microsoft/bitnet-b1.58-2B-4T</a>

WinbuzzerMicrosoft Releases BitNet b1.58 2B4T, a 1.58-Bit AI Model That Runs on Standard CPUs<a href="https://mastodon.social/tags/AI" class="mention hashtag" rel="tag">#AI</a> <a href="https://mastodon.social/tags/Microsoft" class="mention hashtag" rel="tag">#Microsoft</a> <a href="https://mastodon.social/tags/LLMs" class="mention hashtag" rel="tag">#LLMs</a> <a href="https://mastodon.social/tags/BitNet" class="mention hashtag" rel="tag">#BitNet</a> <a href="https://mastodon.social/tags/OpenSource" class="mention hashtag" rel="tag">#OpenSource</a> <a href="https://mastodon.social/tags/MachineLearning" class="mention hashtag" rel="tag">#MachineLearning</a> <a href="https://mastodon.social/tags/CPU" class="mention hashtag" rel="tag">#CPU</a> <a href="https://mastodon.social/tags/DeepLearning" class="mention hashtag" rel="tag">#DeepLearning</a> <a href="https://mastodon.social/tags/1bitLLM" class="mention hashtag" rel="tag">#1bitLLM</a> <a href="https://mastodon.social/tags/GenAI" class="mention hashtag" rel="tag">#GenAI</a><a href="https://winbuzzer.com/2025/04/17/microsoft-releases-bitnet-b1-58-2b4t-a-1-58-bit-ai-model-that-runs-on-standard-cpus-xcxwbn/" target="_blank" rel="nofollow noopener" translate="no">https://winbuzzer.com/2025/04/17/microsoft-releases-bitnet-b1-58-2b4t-a-1-58-bit-ai-model-that-runs-on-standard-cpus-xcxwbn/</a>

N-gated Hacker News🎉🎊 Behold, the groundbreaking 587th iteration of <a href="https://mastodon.social/tags/BitNet" class="mention hashtag" rel="tag">#BitNet</a>, where <a href="https://mastodon.social/tags/buzzwords" class="mention hashtag" rel="tag">#buzzwords</a> meet their ultimate destiny: "Technical Report"! A riveting tale of acronyms, citations, and a plea for <a href="https://mastodon.social/tags/donations" class="mention hashtag" rel="tag">#donations</a>, all while you desperately try to figure out if those numbers actually mean anything 📊🤯. Remember, it's not a real tech report without a job ad and some grateful acknowledgments for funding! 💰👏 <a href="https://arxiv.org/abs/2504.12285" target="_blank" rel="nofollow noopener" translate="no">https://arxiv.org/abs/2504.12285</a> <a href="https://mastodon.social/tags/TechnicalReport" class="mention hashtag" rel="tag">#TechnicalReport</a> <a href="https://mastodon.social/tags/TechNews" class="mention hashtag" rel="tag">#TechNews</a> <a href="https://mastodon.social/tags/HackerNews" class="mention hashtag" rel="tag">#HackerNews</a> <a href="https://mastodon.social/tags/ngated" class="mention hashtag" rel="tag">#ngated</a>

Hacker NewsBitNet b1.58 2B4T Technical Report<a href="https://arxiv.org/abs/2504.12285" target="_blank" rel="nofollow noopener" translate="no">https://arxiv.org/abs/2504.12285</a><a href="https://mastodon.social/tags/HackerNews" class="mention hashtag" rel="tag">#HackerNews</a> <a href="https://mastodon.social/tags/BitNet" class="mention hashtag" rel="tag">#BitNet</a> <a href="https://mastodon.social/tags/b1" class="mention hashtag" rel="tag">#b1</a>.58 <a href="https://mastodon.social/tags/2B4T" class="mention hashtag" rel="tag">#2B4T</a> <a href="https://mastodon.social/tags/Technical" class="mention hashtag" rel="tag">#Technical</a> <a href="https://mastodon.social/tags/Report" class="mention hashtag" rel="tag">#Report</a> <a href="https://mastodon.social/tags/arxiv" class="mention hashtag" rel="tag">#arxiv</a> <a href="https://mastodon.social/tags/research" class="mention hashtag" rel="tag">#research</a> <a href="https://mastodon.social/tags/machinelearning" class="mention hashtag" rel="tag">#machinelearning</a> <a href="https://mastodon.social/tags/AI" class="mention hashtag" rel="tag">#AI</a>

Mr Tech KingMicrosoft open-sourced BitNet b1.58, a big 2B param 1-bit AI. Super efficient on CPUs like M2, it rivals similar models using less memory/speed. Needs a custom framework for now, GPU support pending. <a href="https://mastodon.social/tags/AI" class="mention hashtag" rel="tag">#AI</a> <a href="https://mastodon.social/tags/Microsoft" class="mention hashtag" rel="tag">#Microsoft</a> <a href="https://mastodon.social/tags/BitNet" class="mention hashtag" rel="tag">#BitNet</a>

David Weir :v_enby:Was looking at the source to a very early arXiv paper (<a href="https://arxiv.org/abs/hep-ph/9210243" rel="nofollow noopener" translate="no" target="_blank">https://arxiv.org/abs/hep-ph/9210243</a>). The PDF is unavailable, for reasons that are obscure ("pre-1996 submission which cannot be processed"). But there's a lot of history in the source code: it looks like it was submitted, as a single file, emailed from BITNET to the arXiv via a gateway. It also uses a now-obscure TeX package phyzzx (<a href="https://ctan.org/tex-archive/obsolete/macros/phyzzx" rel="nofollow noopener" translate="no" target="_blank">https://ctan.org/tex-archive/obsolete/macros/phyzzx</a>).I know I'll sound like a young person when I say this but I'd love to know how that worked in practice and what it was like to be in academia before everyone had access to a TCP/IP internet connection but after internetworked computers were ubiquitous. Sort of like the TV series Halt and Catch Fire but with physicists.<a href="https://mementomori.social/tags/arXiv" class="mention hashtag" rel="nofollow noopener" target="_blank">#arXiv</a> <a href="https://mementomori.social/tags/BITNET" class="mention hashtag" rel="nofollow noopener" target="_blank">#BITNET</a> <a href="https://mementomori.social/tags/TeX" class="mention hashtag" rel="nofollow noopener" target="_blank">#TeX</a>

Greg VernonAnyone else here from the <a href="https://mastodon.green/tags/BITNET" class="mention hashtag" rel="nofollow noopener" target="_blank">#BITNET</a> days?

TQWow! 💻 Just discovered BitNet b1.58, a new kind of Large Language Model (LLM). Instead of using standard 16-bit numbers, its weights are simplified to just -1, 0, or 1 (around 1.58 bits)! <a href="https://mastodon.social/tags/LLM" class="mention hashtag" rel="tag">#LLM</a> <a href="https://mastodon.social/tags/AI" class="mention hashtag" rel="tag">#AI</a> <a href="https://mastodon.social/tags/BitNet" class="mention hashtag" rel="tag">#BitNet</a>

G.O.R.NLLMの未来: スケーリング則の限界と効率化の新アプローチ今、LLMは一つの岐路にあると思っている、現状の認識としてはスケーリング則に限界が見受けられること。スケーリング即とはモデルの大規模化によって、モデルの精度、アウトプットの品質が高まるという経験則を指す。しかし、スケーリング即に現状、限界が見えていて、モデルの大規模化が必ずしもアウトプットの深化に結び付かない例が観測されている。<a href="https://note.com/yoshiyuki_hongoh/n/nad718a062020" rel="nofollow noopener" target="_blank">AIの天井が見えてきた日：スケール則の限界と新時代の幕開け</a>最近の事例から、私は今後の有望な方向性として二つのアプローチを見出しています。MoE (Mixture of Experts)：比較的歴史の長い発想で、Gradient Boostingなどもその考え方と考えられます。複数のモデルを組み合わせ、それぞれに重み付けを行うことで高精度なモデルを構築する手法です。BitNet：比較的新しい考え方で、ニューラルネットワークは発火しているか否かという状態を表すため、ビットレンジを極限まで圧縮できるはずという理論に基づきます。このアプローチでは、計算リソースの使用を大幅に削減することが可能です。これまでは、モデルの大規模化競争が主流でしたが、これは妥当な方向性でしょうか。ビッグプレイヤーは結果的に原子力発電にまで進んでいますが、これがよい方向性だとは思えません。したがって、モデルの効率化が今後のゲームチェンジャーとなり得ると考えています。<a rel="nofollow noopener" class="hashtag u-tag u-category" href="https://tech.grayrecord.com/tag/bitnet/" target="_blank">#BitNet</a> <a rel="nofollow noopener" class="hashtag u-tag u-category" href="https://tech.grayrecord.com/tag/llm/" target="_blank">#LLM</a> <a rel="nofollow noopener" class="hashtag u-tag u-category" href="https://tech.grayrecord.com/tag/moe/" target="_blank">#MoE</a> <a rel="nofollow noopener" class="hashtag u-tag u-category" href="https://tech.grayrecord.com/tag/%e3%82%b9%e3%82%b1%e3%83%bc%e3%83%aa%e3%83%b3%e3%82%b0%e5%89%87/" target="_blank">#スケーリング則</a>

michabbb💻 Supports running 100B <a href="https://social.vivaldi.net/tags/BitNet" class="mention hashtag" rel="nofollow noopener" target="_blank">#BitNet</a> b1.58 model on single CPU at 5-7 tokens/sec 🛠️ Built on <a href="https://social.vivaldi.net/tags/opensource" class="mention hashtag" rel="nofollow noopener" target="_blank">#opensource</a> <a href="https://social.vivaldi.net/tags/llamacpp" class="mention hashtag" rel="nofollow noopener" target="_blank">#llamacpp</a> framework with optimized kernels 🔄 Compatible with existing 1-bit models from <a href="https://social.vivaldi.net/tags/HuggingFace" class="mention hashtag" rel="nofollow noopener" target="_blank">#HuggingFace</a> 📱 Future support planned for <a href="https://social.vivaldi.net/tags/NPU" class="mention hashtag" rel="nofollow noopener" target="_blank">#NPU</a> and <a href="https://social.vivaldi.net/tags/GPU" class="mention hashtag" rel="nofollow noopener" target="_blank">#GPU</a> platforms

Tuomas Väisänen 📼🧟‍♂️LLM/AI

frdnd[bitnet HF1BitLLM/Llama3-8B-1.58-100B-tokens -n 128 -t 0]What is a llm? Answer: A llm is a type of essay that is written in the form of a question. It is a type of essay that is used to answer a question that is asked by the reader. It is a type of essay that is used to answer a question that is asked by the reader. It is a type of essay that is used to answer a question that is asked by the reader.Surprisingly fast on CPU but not yet there: <a href="https://github.com/microsoft/BitNet?tab=readme-ov-file" target="_blank" rel="nofollow noopener" translate="no">https://github.com/microsoft/BitNet?tab=readme-ov-file</a> <a href="https://mastodon.social/tags/llm" class="mention hashtag" rel="tag">#llm</a> <a href="https://mastodon.social/tags/bitnet" class="mention hashtag" rel="tag">#bitnet</a>

sumiyakiBitNet b1.58(BitLinear)を実装してMNISTで検証してみた(Tensorflow/Torch) <a href="https://qiita.com/pocokhc/items/09128e92654783a5fa5b" rel="nofollow noopener" target="_blank">https://qiita.com/pocokhc/items/09128e92654783a5fa5b</a> <a href="https://misskey.cloud/tags/AI" rel="nofollow noopener" target="_blank">#AI</a> <a href="https://misskey.cloud/tags/BitNet" rel="nofollow noopener" target="_blank">#BitNet</a> <a href="https://misskey.cloud/tags/MNIST" rel="nofollow noopener" target="_blank">#MNIST</a> <a href="https://misskey.cloud/tags/learning" rel="nofollow noopener" target="_blank">#learning</a>

GM7077I just became aware that vampire capitalists highjacked the <a href="https://masto.ai/tags/BITNET" class="mention hashtag" rel="nofollow noopener" target="_blank">#BITNET</a> of old (in name only, mind you) for a God-forsaken <a href="https://masto.ai/tags/blockchain" class="mention hashtag" rel="nofollow noopener" target="_blank">#blockchain</a> <a href="https://masto.ai/tags/fintech" class="mention hashtag" rel="nofollow noopener" target="_blank">#fintech</a>... Nothing is sacred in this world...<a href="https://masto.ai/tags/IT" class="mention hashtag" rel="nofollow noopener" target="_blank">#IT</a> <a href="https://masto.ai/tags/Tech" class="mention hashtag" rel="nofollow noopener" target="_blank">#Tech</a>

sumiyakiBitNet in training data: 「 Since gradient-based training does not work with 1-bit or binarized networks, non-gradient-based technologies become relevant (checknevergrad and PyGAD), like genetic algorithms or other gradient-free technologies. 」 <a href="https://misskey.cloud/tags/AI" rel="nofollow noopener" target="_blank">#AI</a> <a href="https://misskey.cloud/tags/BitNet" rel="nofollow noopener" target="_blank">#BitNet</a> <a href="https://misskey.cloud/tags/gradient" rel="nofollow noopener" target="_blank">#gradient</a> <a href="https://misskey.cloud/tags/training" rel="nofollow noopener" target="_blank">#training</a>

sumiyakiThe Revolutionary Potential of 1-Bit Language Models (LLMs) <a href="https://hackernoon.com/the-revolutionary-potential-of-1-bit-language-models-llms" rel="nofollow noopener" target="_blank">https://hackernoon.com/the-revolutionary-potential-of-1-bit-language-models-llms</a> <a href="https://misskey.cloud/tags/AI" rel="nofollow noopener" target="_blank">#AI</a> <a href="https://misskey.cloud/tags/BitNet" rel="nofollow noopener" target="_blank">#BitNet</a> <a href="https://misskey.cloud/tags/LLM" rel="nofollow noopener" target="_blank">#LLM</a>

lqdevThe Era of 1-bit LLMs: All Large Language Models are in 1.58 Bits <a href="https://toot.lqdev.tech/tags/llm" class="mention hashtag" rel="nofollow noopener" target="_blank">#llm</a> <a href="https://toot.lqdev.tech/tags/ai" class="mention hashtag" rel="nofollow noopener" target="_blank">#ai</a> <a href="https://toot.lqdev.tech/tags/bitnet" class="mention hashtag" rel="nofollow noopener" target="_blank">#bitnet</a> <a href="https://www.luisquintanilla.me/feed/era-1-bit-llms?utm_medium=feed" rel="nofollow noopener" translate="no" target="_blank">https://www.luisquintanilla.me/feed/era-1-bit-llms?utm_medium=feed</a>

sumiyaki1ビットLLMの衝撃! 70Bで8.9倍高速　全ての推論を加算のみで!GPU不要になる可能性も <a href="https://wirelesswire.jp/2024/02/86094/" rel="nofollow noopener" target="_blank">https://wirelesswire.jp/2024/02/86094/</a> <a href="https://misskey.cloud/tags/AI" rel="nofollow noopener" target="_blank">#AI</a> <a href="https://misskey.cloud/tags/LLM" rel="nofollow noopener" target="_blank">#LLM</a> <a href="https://misskey.cloud/tags/BitNet" rel="nofollow noopener" target="_blank">#BitNet</a> <a href="https://misskey.cloud/tags/BitNetTransformer" rel="nofollow noopener" target="_blank">#BitNetTransformer</a>

GripNews🌗 全面以1.58位元1位元大型語言模型時代 ➤ BitNet：1位元大型語言模型新時代的崛起 ✤ <a href="https://arxiv.org/abs/2402.17764" target="_blank" rel="nofollow noopener" translate="no">https://arxiv.org/abs/2402.17764</a> 最新研究，如BitNet，正在開創一個新的1位元大型語言模型（LLMs）時代。他們提出一種1位元LLM變體，即BitNet b1.58，能在成本效益上明顯優於全精度（例如FP16或BF16）變壓器LLM，同時保持性能和成本效益。1.58位元LLM確立了訓練新世代高性能、具成本效益LLMs的新法則和配方，同時開啟了針對1位元LLMs進行優化的特定硬體設計新範式。 + 對於語言模型技術的最新發展感到興奮，1位元模型的概念令人印象深刻。 + 1.58位元LLM將有望帶來訓練成本和效能之間更好的平衡，是未來發展的有趣方向。 <a href="https://mastodon.social/tags/%E5%A4%A7%E5%9E%8B%E8%AA%9E%E8%A8%80%E6%A8%A1%E5%9E%8B" class="mention hashtag" rel="tag">#大型語言模型</a> <a href="https://mastodon.social/tags/1%E4%BD%8D%E5%85%83%E6%A8%A1%E5%9E%8B" class="mention hashtag" rel="tag">#1位元模型</a> <a href="https://mastodon.social/tags/BitNet" class="mention hashtag" rel="tag">#BitNet</a>

awbBefore I could access the Internet, I was on BITNET. For a system based on remote job entry, it was surprisingly useful and fun! LISTSERV originated there and most communication was non-interactive. There still was a way to do messaging and even a chat network existed that inspired IRC. The limitations of BITNET resulted in a lot of creative solutions. <a href="https://mastodon.sdf.org/tags/bitnet" class="mention hashtag" rel="nofollow noopener" target="_blank">#bitnet</a> <a href="https://mastodon.sdf.org/tags/listserv" class="mention hashtag" rel="nofollow noopener" target="_blank">#listserv</a>

Recent searches

Search options

Administered by:

Server stats:

#bitnet