@kyutai_labs 在 Huggingface 以及 GitHub 上釋出了即時語音對話模型，該模型是個 7B 模型，如果你有張 RTX 3090 24GB 或者與之大小相符的顯卡就可以跑這顆 Model 了，再來就可能要選擇量化模型了。

相信大家可能或多或少都有看到這篇在 Ｘ 上的貼文，這是一個來自法國的初創團隊 Kyutai，在七月初發佈對標 GPT-4o 的神奇模型。經歷兩個多月的時間，他們也對兌現自己的承諾，不僅把模型開源，甚至程式碼也開源了！

論文: https://kyutai.org/Moshi.pdf
開源程式碼: https://github.com/kyutai-labs/moshi
開源模型: https://huggingface.co/collections/kyutai/moshi-…
線上體驗: moshi.chat

大家可以去官網體驗看看 (moshi.chat)，能夠在落地的環境採用這些像人一般的模型，還是很奇妙的。

甚至我心目中的 AI 英雄 Andrej Karpathy 也在 X 上給出了讚賞

技術細節

• 模型參數量: 7.69B

• 支援平台: PyTorch、Candle、MLX

• 語音輸入處理: Mimi; a neural audio codec (paper:3)

• ALM 架構: 結合 Helium 和較小的 Transformer 模型

模型的參數量約為 7.96B，模型支援以下三種平台 PyTorch、MLX(An array framework for Apple silicon) 以及 Candle (Minimalist ML framework for Rust)，如果你想執行 Quantized 的模型，則 Candle 提供了 8bit 的版本，MLX 則還有 4bit 的可以使用。

Python (MLX) for local inference on macOS

Installation

假如你想在自己電腦上執行 moshi，以下是 Mac 的快速安裝。

你至少需要 Python 3.10，建議使用 3.12。有關具體需求，請查看個別狀況而定。
你可以使用下列方式安裝 PyTorch 和 MLX 用戶端：

pip install moshi      # moshi PyTorch, from PyPI
pip install moshi_mlx  # moshi MLX, from PyPI, best with Python 3.12.
# Or the bleeding edge versions for Moshi and Moshi-MLX.
pip install -e "git+https://git@github.com/kyutai-labs/moshi.git#egg=moshi&subdirectory=moshi"
pip install -e "git+https://git@github.com/kyutai-labs/moshi.git#egg=moshi_mlx&subdirectory=moshi_mlx"

pip install rustymimi  # mimi, rust implementation with Python bindings from PyPI

local inference

python -m moshi_mlx.local -q 4   # 執行 4bit 模型;weights quantized to 4 bits
python -m moshi_mlx.local -q 8   # 執行 8bit 模型;weights quantized to 8 bits
# 使用其他不同的預訓練模型:
python -m moshi_mlx.local -q 4 --hf-repo kyutai/moshika-mlx-q4
python -m moshi_mlx.local -q 8 --hf-repo kyutai/moshika-mlx-q8
# 請注意要配合使用 "-q" 以及 "--hf-repo"

如果想了解 Moshi 更細緻的技術細節，可以透過下方訂閱我的 News letter 即時追蹤 AI 產業的最新趨勢。

What if you're looking for a new blog platform for various purposes, whether it's for development logs, personal growth, or more? Here's a beginner's guide to using Hashnode as your blogging platform.

I want to guide everyone through registering, setting up a custom domain, and exploring the AI features offered. It's a simple, step-by-step approach to quickly getting your blog up and running with these AI tools.

Registration

To register, you can only go to https://hashnode.com/ and click Sign up for free.

You can choose from several options to sign up. They support GitHub, Apple ID, LinkedIn, Facebook, Google, and even SSO on your custom domain.

If you prefer not to use those accounts, you can also sign up using just your email.

Create your first blog with just one click

At the top, you'll see a New Blog button. Click it, enter your blog name and the URL you want, and then hit Create. That's it — you're all set!

Writing Your First Post

You can see the "New Article" button on the top right to create a new article and start writing.

AI-Powered features

Forget about ChatGPT;

Here, you just need to Select and Click. I've never used Generative AI that's this fast and easy.

Conclusion

Embarking on your blogging journey with Hashnode is both easy and rewarding. From a seamless registration process using your preferred sign-up method to effortlessly creating your first blog with just a few clicks, Hashnode makes it simple to get started. Once your blog is up and running, you can dive right into writing engaging content using the intuitive "New Article" feature.

The AI-powered tools available on Hashnode elevate your writing experience, allowing you to generate ideas, enhance your drafts, and maintain a consistent publishing schedule with minimal effort. Whether you're sharing development insights, personal growth stories, or exploring other topics, Hashnode provides the essential tools and flexibility to help your blog thrive.

Stay tuned for our next post, where we'll guide you through setting up a custom domain to further personalize your blog and establish your unique online presence. Until then, take advantage of the powerful features Hashnode offers and start creating content that resonates with your audience. Happy blogging!