China just

@dr_cintas

57.3K followers · usually 4.0K views a post

Views

99.2K

Likes

7.0K

Comments

345

1.0K

SoundTikTok sound · original sound - dr_cintas

Promoted product

Kimi K3

Why it worked

The video effectively uses a fast-paced, visually engaging format to deliver a high volume of AI news in a short time. The inclusion of diverse AI tools and models, presented with clear on-screen text and concise explanations, caters to an audience interested in staying updated on AI advancements.

Summary

This video provides a weekly roundup of AI news, highlighting several new tools and models. It features Kimi K3, a portable AI model that runs on CPU, Pocket TTS for voice cloning, Hermes Skills Hub for AI agents, Grok Imagine for image editing, Muse Code as a coding agent, and new models from Qwen and Meta. The video also touches on advancements in AI image generation and large language models.

Structure

1Introduction to AI news roundup
2Highlighting Kimi K3 and Pocket TTS
3Showcasing Hermes Skills Hub and its capabilities
4Demonstrating Grok Imagine for image editing
5Introducing Muse Code and its function
6Announcing new AI models from Qwen and Meta

Product placement

Kimi K3 (a 2.78-trillion-parameter AI model that runs inference on a single CPU with 8.24 GB of RAM, and is portable C99 with no BLAS, no framework, and no GPU) ACTS - it is the subject of the video, with its capabilities and performance discussed. Removing it would change the video. Pocket TTS (a text-to-speech model with 100M parameters that runs on CPU, has low latency, and can clone voices) ACTS - it is presented as a new tool that can be run on a user's own computer. Removing it would change the video. Hermes Skills Hub (a platform to discover, search, and install AI skills) ACTS - it is shown as a resource for AI agents, with over 90,000 skills across 200 categories. Removing it would change the video. OpenAI (a company that develops AI models) APPEARS - its logo is shown as one of the providers of AI skills. Removing it would not change the video. Anthropic (a company that develops AI models) APPEARS - its logo is shown as one of the providers of AI skills. Removing it would not change the video. NVIDIA (a company that develops GPUs and AI technology) APPEARS - its logo is shown as one of the providers of AI skills. Removing it would not change the video. Grok Imagine Image 2.0 (an AI image generation tool) ACTS - it is demonstrated as a tool that allows users to change specific regions of an image and its background. Removing it would change the video. Muse Code (Meta's first coding agent) ACTS - it is presented as a tool that runs in the terminal and keeps background agents alive. Removing it would change the video. Kimi K3 (mentioned again, referring to its pure C99 engine) ACTS - its engine is described as streaming experts from disk instead of loading them into memory. Removing it would change the video. Qwen3.8-Max (an AI model with 2.4T parameters, 95B active) ACTS - it is presented as a new model available via QwenCloud, with improvements across various tasks. Removing it would change the video. Simplifying AI (a website offering AI and tech news) APPEARS - it is the creator's bio link destination. Removing it would not change the video.

Original caption

Here is everything that happened in AI this week in 60 seconds 🤯🚨

More from @dr_cintas

On-screen text

CHINA China just most advanced open source while this new tool let's run a parameter model ON A LAPTOP 8 GIGS OF 8 GIGS OF RAM AND NO G all that happened in AI this week in seconds 6 /kyutai Pocket TTS Supports Python 3.10, 3.11, 3.12, 3.13 and 3.14. Requires PyTorch 2.5+. Does not require the gpu version of PyTorch. Demo | GitHub Repository | Hugging Face Model Card | Tech report | Paper | Documentation Main takeaways Runs on CPU Small model size, 100M parameters Audio streaming Low latency, ~200ms to get the first audio chunk Faster than real-time, ~6x real-time on a CPU of MacBook Air M4 Uses only 2 CPU cores Python API and CLI Voice cloning Multi-language support: english, french, german, portuguese, italian, spanish Can handle infinitely long text inputs Can run on client-side in the browser Additional languages can be added in the future. Using it as a Python library You can try out the Python library on Colab here. Install the package with pip install pocket-tts # or uv add pocket-tts is literally one command 5 reference NOU HERMES SKILLS HUB over 90,000 for your to work across 200 200 CATEGORIES OpenAI ANTHROPIC OpenAI NVIDIA installable into your agent 4 Grok Imagine IMAGE 2.0 A PRECISE TOOL FOR THE IMAGINATION and change plus background 5 reference smart ratio smart resize It's now NO. 2 IN THE WORLD OF IMAGE EDITING 3 Muse Code. Muse Code it's 1st coding agent terminal background across session in parallel your repo Qwen3.8 Qwen3.8-Max Qwen3.8-Max - now available via QwenCloud: 2.4T parameters (95B active), with open weights releasing next week comprehensive improvements across coding, work, research, and long-horizon tasks end-to-end and dependable delivery of complex tasks Call via API on QwenCloud. 2.4 trillion parameters only 95 billion 1 Explore the Epic Story of Life Start Exploring 1 MILLION TOKEN CONTEXT WINDOW native video understanding the highest Chinese model 1 Kimi K3 K3 pulsar research deck with P-Pdot chart and report. kimik3 kimik3-in-c A 2.78-trillion-parameter model. One CPU. 8 GB of RAM. Kimi K3 inference in portable C99. No BLAS. No framework. No GPU. CI passing license Apache 2.0 C99 portable platform Linux x86-64 version 1.0.0 2.78T 1.56 TB checkpoint on disk 8.24 GB peak RSS, measured 176 KB the whole engine KIMI K3 ENGINE MIXTURE-OF-EXPERTS SPARSE ROUTING 16 of 896 fire CPU ONLY FROM SCRATCH IN C FIXED-SIZE STATE context costs nothing extra STREAM THE TRUNK a dial, not a floor RESIDENT IN RAM (TOTAL: 8.24 GB) LM HEAD 163,840 logits ARGMAX (Next Token) DETOKENIZER (BPE -> Text) OUTPUT "Paris." MLA KV CACHE 2.37 MB per position (Latent) EXPERT LRU CACHE 17.56 MB slots (Hot Experts) EXPERT PAYLOAD 25.58 GB per token, 1,472 reads TOTAL ~135 GB / TOKEN BYPASSED PAGE CACHE Direct I/O O_DIRECT No Page Cache EXPERT POOL 896 per layer, 82,432 total, 1.45 TB OFFLINE PREPARATION (ONE-TIME) BUILD PINNED TRUNK PACK & INDEX EXPERTS WRITE SHARDS TO NVMe Crossing Gates Storage Offline Reductions Disabled kimik3 of your disk loading them Generate Assets into memory Paris. The first token out of a 2.78 trillion parameter model running on a CPU in a few gigabytes of RAM is the right answer. The trailing quote and the "The Eiffel" continuation are not a defect: with no chat template a no instruction tuning in the loop, this is a base model continuing text. It has decided it is inside a JSON list of sentences about France, and it is continuing that list. with no 100% 100% OPEN SOURCED them out? AI

Transcript

China just most advanced open source while this new tool let's run a parameter model ON A LAPTOP 8 GIGS OF 8 GIGS OF RAM AND NO G all that happened in AI this week in seconds 6 /kyutai Pocket TTS Supports Python 3.10, 3.11, 3.12, 3.13 and 3.14. Requires PyTorch 2.5+. Does not require the gpu version of PyTorch. Demo | GitHub Repository | Hugging Face Model Card | Tech report | Paper | Documentation Main takeaways Runs on CPU Small model size, 100M parameters Audio streaming Low latency, ~200ms to get the first audio chunk Faster than real-time, ~6x real-time on a CPU of MacBook Air M4 Uses only 2 CPU cores Python API and CLI Voice cloning Multi-language support: english, french, german, portuguese, italian, spanish Can handle infinitely long text inputs Can run on client-side in the browser Additional languages can be added in the future. Using it as a Python library You can try out the Python library on Colab here. Install the package with pip install pocket-tts # or uv add pocket-tts is literally one command 5 reference NOU HERMES SKILLS HUB over 90,000 for your to work across 200 200 CATEGORIES OpenAI ANTHROPIC OpenAI NVIDIA installable into your agent 4 Grok Imagine IMAGE 2.0 A PRECISE TOOL FOR THE IMAGINATION and change plus background 5 reference smart ratio smart resize It's now NO. 2 IN THE WORLD OF IMAGE EDITING 3 Muse Code. Muse Code it's 1st coding agent terminal background across session in parallel your repo Qwen3.8 Qwen3.8-Max Qwen3.8-Max - now available via QwenCloud: 2.4T parameters (95B active), with open weights releasing next week comprehensive improvements across coding, work, research, and long-horizon tasks end-to-end and dependable delivery of complex tasks Call via API on QwenCloud. 2.4 trillion parameters only 95 billion 1 Explore the Epic Story of Life Start Exploring 1 MILLION TOKEN CONTEXT WINDOW native video understanding the highest Chinese model 1 Kimi K3 K3 pulsar research deck with P-Pdot chart and report. kimik3 kimik3-in-c A 2.78-trillion-parameter model. One CPU. 8 GB of RAM. Kimi K3 inference in portable C99. No BLAS. No framework. No GPU. CI passing license Apache 2.0 C99 portable platform Linux x86-64 version 1.0.0 2.78T 1.56 TB checkpoint on disk 8.24 GB peak RSS, measured 176 KB the whole engine KIMI K3 ENGINE MIXTURE-OF-EXPERTS SPARSE ROUTING 16 of 896 fire CPU ONLY FROM SCRATCH IN C FIXED-SIZE STATE context costs nothing extra STREAM THE TRUNK a dial, not a floor RESIDENT IN RAM (TOTAL: 8.24 GB) LM HEAD 163,840 logits ARGMAX (Next Token) DETOKENIZER (BPE -> Text) OUTPUT "Paris." MLA KV CACHE 2.37 MB per position (Latent) EXPERT LRU CACHE 17.56 MB slots (Hot Experts) EXPERT PAYLOAD 25.58 GB per token, 1,472 reads TOTAL ~135 GB / TOKEN BYPASSED PAGE CACHE Direct I/O O_DIRECT No Page Cache EXPERT POOL 896 per layer, 82,432 total, 1.45 TB OFFLINE PREPARATION (ONE-TIME) BUILD PINNED TRUNK PACK & INDEX EXPERTS WRITE SHARDS TO NVMe Crossing Gates Storage Offline Reductions Disabled kimik3 of your disk loading them Generate Assets into memory Paris. The first token out of a 2.78 trillion parameter model running on a CPU in a few gigabytes of RAM is the right answer. The trailing quote and the "The Eiffel" continuation are not a defect: with no chat template a no instruction tuning in the loop, this is a base model continuing text. It has decided it is inside a JSON list of sentences about France, and it is continuing that list. with no 100% 100% OPEN SOURCED them out? AI