On the Shifting Global Compute Landscape
โข
51
AI & ML interests
The AI community building the future.
Recent Activity
View all activity
Papers
View all Papers
Articles
sayakpaulย
updated
a
dataset
about 4 hours ago
alvarobarttย
updated
a
dataset
about 5 hours ago
nielsrย
updated
a
Space
about 8 hours ago
lysandreย
updated
a
dataset
about 9 hours ago
eustlbย
authored
a
paper
11 days ago
Steveeeeeeenย
authored
2
papers
15 days ago
Post
355
Building high-performance, reproducible kernels forย AMD ROCmย just got a lot easier.
I've put together a guide on building, testing, and sharingย ROCm-compatible kernelsย using the Hugging Faceย
Learn how to:
- Use Nix for reproducible builds
- Integrate kernels as native PyTorch operators
- Share your kernels on the Hub for anyone to use withย
We use the ๐ award-winning RadeonFlow GEMM kernel as a practical example.
๐ Check out the full guide here : https://huggingface.co/blog/build-rocm-kernels
I've put together a guide on building, testing, and sharingย ROCm-compatible kernelsย using the Hugging Faceย
kernel-builder and kernelsย libraries; so you can focus onย optimizing performanceย rather than spending time on setup.Learn how to:
- Use Nix for reproducible builds
- Integrate kernels as native PyTorch operators
- Share your kernels on the Hub for anyone to use withย
kernels.get_kernel()We use the ๐ award-winning RadeonFlow GEMM kernel as a practical example.
๐ Check out the full guide here : https://huggingface.co/blog/build-rocm-kernels
Post
2181
Hugging Face MCP Server v0.2.46
~~~~~~~~~~~~~~~~~~~~~~~~~~~~
- Add "discover" to Dynamic Space tool. Recommend deselecting "space_search" if using dynamic spaces.
~~~~~~~~~~~~~~~~~~~~~~~~~~~~
- Add "discover" to Dynamic Space tool. Recommend deselecting "space_search" if using dynamic spaces.
Post
2887
Hugging Face MCP Server v0.2.45
~~~~~~~~~~~~~~~~~~~~~~~~~~~~
- New! Experimental dynamic_space tool.
- Default Image Generator changed to Qwen-Image-Fast
~~~~~~~~~~~~~~~~~~~~~~~~~~~~
- New! Experimental dynamic_space tool.
- Default Image Generator changed to Qwen-Image-Fast
Post
2147
Hugging Face MCP Server v0.2.40
~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Improved progressive disclosure and descriptions for Jobs tool.
~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Improved progressive disclosure and descriptions for Jobs tool.
Post
3194
Kimi K2 Thinking is now live on the hub ๐ฅ
moonshotai/Kimi-K2-Thinking
โจ 1T MoE for deep reasoning & tool use
โจ Native INT4 quantization = 2ร faster inference
โจ 256K context window
โจ Modified MIT license
moonshotai/Kimi-K2-Thinking
โจ 1T MoE for deep reasoning & tool use
โจ Native INT4 quantization = 2ร faster inference
โจ 256K context window
โจ Modified MIT license
Post
627
Chinese open source AI in October wasnโt about bigger models, it was about real world impact ๐ฅ
https://huggingface.co/collections/zh-ai-community/october-2025-china-open-source-highlights
โจ Vision-Language & OCR wave ๐
- DeepSeek-OCR : 3B
- PaddleOCR-VL : 0.9B
- Qwen3-VL : 2B / 4B / 8B / 32B /30B-A3B
- Open-Bee: Bee-8B-RL
- http://Z.ai Glyph :10B
OCR is industrializing, the real game now is understanding the (long context) document, not just reading it.
โจ Text generation: scale or innovation?
- MiniMax-M2: 229B
- Antgroup Ling-1T & Ring-1T
- Moonshot Kimi-Linear : linear-attention challenger
- Kwaipilot KAT-Dev
Efficiency is the key.
โจ Any-to-Any & World-Model : one step forward to the real world
- BAAI Emu 3.5
- Antgroup Ming-flash-omni
- HunyuanWorld-Mirror: 3D
Aligning with the โworld modelโ globally
โจ Audio & Speech + Video & Visual: released from entertainment labs to delivery platforms
- SoulX-Podcast TTS
- LongCat-Audio-Codec & LongCat-Video by Meituan delivery paltform
- xiabs DreamOmni 2
Looking forward to what's next ๐
https://huggingface.co/collections/zh-ai-community/october-2025-china-open-source-highlights
โจ Vision-Language & OCR wave ๐
- DeepSeek-OCR : 3B
- PaddleOCR-VL : 0.9B
- Qwen3-VL : 2B / 4B / 8B / 32B /30B-A3B
- Open-Bee: Bee-8B-RL
- http://Z.ai Glyph :10B
OCR is industrializing, the real game now is understanding the (long context) document, not just reading it.
โจ Text generation: scale or innovation?
- MiniMax-M2: 229B
- Antgroup Ling-1T & Ring-1T
- Moonshot Kimi-Linear : linear-attention challenger
- Kwaipilot KAT-Dev
Efficiency is the key.
โจ Any-to-Any & World-Model : one step forward to the real world
- BAAI Emu 3.5
- Antgroup Ming-flash-omni
- HunyuanWorld-Mirror: 3D
Aligning with the โworld modelโ globally
โจ Audio & Speech + Video & Visual: released from entertainment labs to delivery platforms
- SoulX-Podcast TTS
- LongCat-Audio-Codec & LongCat-Video by Meituan delivery paltform
- xiabs DreamOmni 2
Looking forward to what's next ๐
AdinaYย
authored
a
paper
about 1 month ago
Post
504
Kimi Linear๐ Hybrid linear attention model from Moonshot AI
https://huggingface.co/collections/moonshotai/kimi-linear-a3b
โจ 48B total/ 3B active - MIT license
โจ Up to 1M context
โจ 84.3 on RULER (128k) with 3.98ร speedup
โจ Hybrid KDA + MLA architecture for peak throughput & quality
https://huggingface.co/collections/moonshotai/kimi-linear-a3b
โจ 48B total/ 3B active - MIT license
โจ Up to 1M context
โจ 84.3 on RULER (128k) with 3.98ร speedup
โจ Hybrid KDA + MLA architecture for peak throughput & quality
nouamanetaziย
posted
an
update
about 1 month ago
Post
3942
After training ๐๐ฆ๐จ๐ฅ๐๐๐ on ๐๐๐ ๐๐๐๐๐ฌ for nearly a month, I've come to realize something most people overlook: ๐ข๐ง๐๐ซ๐๐ฌ๐ญ๐ซ๐ฎ๐๐ญ๐ฎ๐ซ๐ ๐ข๐ฌ ๐ญ๐ก๐ ๐ฆ๐๐ค๐-๐จ๐ซ-๐๐ซ๐๐๐ค ๐๐๐๐ญ๐จ๐ซ ๐ข๐ง ๐๐๐ ๐ญ๐ซ๐๐ข๐ง๐ข๐ง๐ . ๐ฅ
Everyone talks about model architecture and data quality. And yes, those matter immensely. But here's what nobody tells you: when your training run fails at 2 AM because of mysterious ๐๐๐๐ ๐๐ซ๐ซ๐จ๐ซ๐ฌ, or when your expensive GPU cluster is running at ๐๐% ๐๐๐๐ข๐๐ข๐๐ง๐๐ฒ, the problem isn't your model. It's most probably a ๐ฆ๐ข๐ฌ๐ฎ๐ฌ๐ ๐จ๐ ๐ญ๐ก๐ ๐ก๐๐ซ๐๐ฐ๐๐ซ๐. ๐ ๏ธ
Questions that seemed simple but had no clear answers: Why is ๐๐จ๐ ๐ญ๐ซ๐๐ข๐ง๐ข๐ง๐ ๐ฌ๐ฅ๐จ๐ฐ๐๐ซ ๐ญ๐ก๐๐ง ๐๐๐ง๐ฌ๐ ๐ฆ๐จ๐๐๐ฅ๐ฌ? Which ๐๐๐๐ ๐๐ฅ๐๐ ๐ฌ should we actually set? How often should we checkpoint without killing throughput?
That's why we built ๐๐ก๐ ๐๐ฆ๐จ๐ฅ ๐๐ซ๐๐ข๐ง๐ข๐ง๐ ๐๐ฅ๐๐ฒ๐๐จ๐จ๐ค ๐: a complete guide covering everything from model architecture and data curation to the SmolLM3 training marathon, post-training techniques, and crucially, the ๐ข๐ง๐๐ซ๐๐ฌ๐ญ๐ซ๐ฎ๐๐ญ๐ฎ๐ซ๐ ๐ฅ๐๐ฒ๐๐ซ that most teams get wrong.
We validated real vs theoretical bandwidth across the entire stack: ๐๐๐๐ ๐ก๐ข๐ญ๐ญ๐ข๐ง๐ ๐ ๐๐/๐ฌ, ๐๐๐๐ข๐ง๐ค ๐.๐ ๐ซ๐๐๐๐ก๐ข๐ง๐ ๐๐๐ ๐๐/๐ฌ, ๐๐๐๐ ๐๐๐ง๐ ๐๐ญ ๐๐.๐ ๐๐/๐ฌ. Then we ran collective operations across ๐๐๐ ๐๐๐๐ฌ (16 nodes, 8xH100s each) and measured how performance degrades at scale: all-reduce drops from ๐๐๐ ๐๐/๐ฌ on a single node to ๐๐๐-๐๐๐ ๐๐/๐ฌ across 16 nodes.
If you've ever wondered why your training runs are slower than they should be, or you're planning to scale up and want to avoid expensive mistakes, this guide might save you weeks of debugging.
๐๐ก๐ ๐๐ฆ๐จ๐ฅ ๐๐ซ๐๐ข๐ง๐ข๐ง๐ ๐๐ฅ๐๐ฒ๐๐จ๐จ๐ค: https://lnkd.in/e5MKXUHS
Shared with โค๏ธ by the HuggingFace team
Everyone talks about model architecture and data quality. And yes, those matter immensely. But here's what nobody tells you: when your training run fails at 2 AM because of mysterious ๐๐๐๐ ๐๐ซ๐ซ๐จ๐ซ๐ฌ, or when your expensive GPU cluster is running at ๐๐% ๐๐๐๐ข๐๐ข๐๐ง๐๐ฒ, the problem isn't your model. It's most probably a ๐ฆ๐ข๐ฌ๐ฎ๐ฌ๐ ๐จ๐ ๐ญ๐ก๐ ๐ก๐๐ซ๐๐ฐ๐๐ซ๐. ๐ ๏ธ
Questions that seemed simple but had no clear answers: Why is ๐๐จ๐ ๐ญ๐ซ๐๐ข๐ง๐ข๐ง๐ ๐ฌ๐ฅ๐จ๐ฐ๐๐ซ ๐ญ๐ก๐๐ง ๐๐๐ง๐ฌ๐ ๐ฆ๐จ๐๐๐ฅ๐ฌ? Which ๐๐๐๐ ๐๐ฅ๐๐ ๐ฌ should we actually set? How often should we checkpoint without killing throughput?
That's why we built ๐๐ก๐ ๐๐ฆ๐จ๐ฅ ๐๐ซ๐๐ข๐ง๐ข๐ง๐ ๐๐ฅ๐๐ฒ๐๐จ๐จ๐ค ๐: a complete guide covering everything from model architecture and data curation to the SmolLM3 training marathon, post-training techniques, and crucially, the ๐ข๐ง๐๐ซ๐๐ฌ๐ญ๐ซ๐ฎ๐๐ญ๐ฎ๐ซ๐ ๐ฅ๐๐ฒ๐๐ซ that most teams get wrong.
We validated real vs theoretical bandwidth across the entire stack: ๐๐๐๐ ๐ก๐ข๐ญ๐ญ๐ข๐ง๐ ๐ ๐๐/๐ฌ, ๐๐๐๐ข๐ง๐ค ๐.๐ ๐ซ๐๐๐๐ก๐ข๐ง๐ ๐๐๐ ๐๐/๐ฌ, ๐๐๐๐ ๐๐๐ง๐ ๐๐ญ ๐๐.๐ ๐๐/๐ฌ. Then we ran collective operations across ๐๐๐ ๐๐๐๐ฌ (16 nodes, 8xH100s each) and measured how performance degrades at scale: all-reduce drops from ๐๐๐ ๐๐/๐ฌ on a single node to ๐๐๐-๐๐๐ ๐๐/๐ฌ across 16 nodes.
If you've ever wondered why your training runs are slower than they should be, or you're planning to scale up and want to avoid expensive mistakes, this guide might save you weeks of debugging.
๐๐ก๐ ๐๐ฆ๐จ๐ฅ ๐๐ซ๐๐ข๐ง๐ข๐ง๐ ๐๐ฅ๐๐ฒ๐๐จ๐จ๐ค: https://lnkd.in/e5MKXUHS
Shared with โค๏ธ by the HuggingFace team
Post
3726
๐ค Did you know your voice might be cloned without your consent from just *one sentence* of audio?
That's not great. So with @frimelle , we brainstormed a new idea for developers who want to curb malicious use: โจThe Voice Consent Gate.โจ
Details, code, here: https://huggingface.co/blog/voice-consent-gate
That's not great. So with @frimelle , we brainstormed a new idea for developers who want to curb malicious use: โจThe Voice Consent Gate.โจ
Details, code, here: https://huggingface.co/blog/voice-consent-gate
- 3 replies
Post
323
Hugging Face MCP Server v0.2.35
~~~~~~~~~~~~~~~~~~~~~~~~~~~~
$HF_TOKEN is expanded in Jobs Secrets environment variables.
~~~~~~~~~~~~~~~~~~~~~~~~~~~~
$HF_TOKEN is expanded in Jobs Secrets environment variables.
Post
1747
Ming-flash-omni Preview ๐ Multimodal foundation model from AntGroup
inclusionAI/Ming-flash-omni-Preview
โจ Built on Ling-Flash-2.0: 10B total/6B active
โจ Generative segmentation-as-editing
โจ SOTA contextual & dialect ASR
โจ High-fidelity image generation
inclusionAI/Ming-flash-omni-Preview
โจ Built on Ling-Flash-2.0: 10B total/6B active
โจ Generative segmentation-as-editing
โจ SOTA contextual & dialect ASR
โจ High-fidelity image generation
Post
1855
Glyph ๐ฅ a framework that scales context length by compressing text into images and processing them with visionโlanguage models, released by Z.ai.
Paper:https://huggingface.co/papers/2510.17800
Model:https://huggingface.co/zai-org/Glyph
โจ Compresses long sequences visually to bypass token limits
โจ Reduces computational and memory costs
โจ Preserves meaning through multimodal encoding
โจ Built on GLM-4.1V-9B-Base