SmolLM2: When Smol Goes Big -- Data-Centric Training of a Small Language Model
Paper
โข
2502.02737
โข
Published
โข
46
Hugging Face Science
company
AI & ML interests
None defined yet.
Recent Activity
View all activity
science's activity
lvwerraย
authored
a
paper
about 3 hours ago
Post
5365
Introducing ๐ผ๐ฝ๐ฒ๐ป ๐๐ฒ๐ฒ๐ฝ-๐ฅ๐ฒ๐๐ฒ๐ฎ๐ฟ๐ฐ๐ต by Hugging Face! ๐ฅ
OpenAI's latest agentic app Deep Research seems really good... But it's closed, as usual.
โฑ๏ธ So with a team of cracked colleagues, we set ourselves a 24hours deadline to replicate and open-source Deep Research! โฑ๏ธ
โก๏ธ We built open-Deep-Research, an entirely open agent that can: navigate the web autonomously, scroll and search through pages, download and manipulate files, run calculation on data...
We aimed for the best performance: are the agent's answers really rigorous?
On GAIA benchmark, Deep Research had 67% accuracy on the validation set.
โก๏ธ open Deep Research is at 55% (powered by o1), it is:
- the best pass@1 solution submitted
- the best open solution ๐ช๐ช
And it's only getting started ! Please jump in, drop PRs, and let's bring it to the top !
Read the blog post ๐ https://huggingface.co/blog/open-deep-research
OpenAI's latest agentic app Deep Research seems really good... But it's closed, as usual.
โฑ๏ธ So with a team of cracked colleagues, we set ourselves a 24hours deadline to replicate and open-source Deep Research! โฑ๏ธ
โก๏ธ We built open-Deep-Research, an entirely open agent that can: navigate the web autonomously, scroll and search through pages, download and manipulate files, run calculation on data...
We aimed for the best performance: are the agent's answers really rigorous?
On GAIA benchmark, Deep Research had 67% accuracy on the validation set.
โก๏ธ open Deep Research is at 55% (powered by o1), it is:
- the best pass@1 solution submitted
- the best open solution ๐ช๐ช
And it's only getting started ! Please jump in, drop PRs, and let's bring it to the top !
Read the blog post ๐ https://huggingface.co/blog/open-deep-research
Post
2280
๐ R1 just built its own download dashboard!
Some fresh stats: +6M downloads for 800+ derivative models vs 2M for originals. Watch the numbers grow here: fdaudens/deepseek-download-stats
Some fresh stats: +6M downloads for 800+ derivative models vs 2M for originals. Watch the numbers grow here: fdaudens/deepseek-download-stats
Post
2433
Now you can launch a code agent directly from your terminal!
โจ ๐๐๐๐๐๐๐๐๐ "๐๐๐๐ ๐๐๐๐" directly launches a CodeAgent
โถ๏ธ This also works with web agents (replace ๐๐๐๐๐๐๐๐๐ with ๐ ๐๐๐๐๐๐๐) thanks to @merve !
๐พ Another treat from smolagents release 1.7.0:
Now agents have a memory mechanism, enabling many possibilities like replaying the last run with ๐๐๐๐๐.๐๐๐๐๐๐ข(), thank you @clefourrier !
Check the release notes here ๐ https://github.com/huggingface/smolagents/releases/tag/v1.7.0
โจ ๐๐๐๐๐๐๐๐๐ "๐๐๐๐ ๐๐๐๐" directly launches a CodeAgent
โถ๏ธ This also works with web agents (replace ๐๐๐๐๐๐๐๐๐ with ๐ ๐๐๐๐๐๐๐) thanks to @merve !
๐พ Another treat from smolagents release 1.7.0:
Now agents have a memory mechanism, enabling many possibilities like replaying the last run with ๐๐๐๐๐.๐๐๐๐๐๐ข(), thank you @clefourrier !
Check the release notes here ๐ https://github.com/huggingface/smolagents/releases/tag/v1.7.0
Post
3204
๐ฏ Kokoro TTS just hit v1.0! ๐
Small but mighty: 82M parameters, runs locally, speaks multiple languages. The best part? It's Apache 2.0 licensed!
This could unlock so many possibilities โจ
Check it out: hexgrad/Kokoro-82M
Small but mighty: 82M parameters, runs locally, speaks multiple languages. The best part? It's Apache 2.0 licensed!
This could unlock so many possibilities โจ
Check it out: hexgrad/Kokoro-82M
- 1 reply
Post
1665
๐ The open source community is unstoppable: 4M total downloads for DeepSeek models on Hugging Face, with 3.2M coming from the +600 models created by the community.
That's 30% more than yesterday!
That's 30% more than yesterday!
- 1 reply
Post
3549
๐ง๐ต๐ฒ ๐๐๐ฏ ๐๐ฒ๐น๐ฐ๐ผ๐บ๐ฒ๐ ๐ฒ๐
๐๐ฒ๐ฟ๐ป๐ฎ๐น ๐ถ๐ป๐ณ๐ฒ๐ฟ๐ฒ๐ป๐ฐ๐ฒ ๐ฝ๐ฟ๐ผ๐๐ถ๐ฑ๐ฒ๐ฟ๐!
โ Hosting our own inference was not enough: now the Hub 4 new inference providers: fal, Replicate, SambaNova Systems, & Together AI.
Check model cards on the Hub: you can now, in 1 click, use inference from various providers (cf video demo)
Their inference can also be used through our Inference API client. There, you can use either your custom provider key, or your HF token, then billing will be handled directly on your HF account, as a way to centralize all expenses.
๐ธ Also, PRO users get 2$ inference credits per month!
Read more in the announcement ๐ https://huggingface.co/blog/inference-providers
โ Hosting our own inference was not enough: now the Hub 4 new inference providers: fal, Replicate, SambaNova Systems, & Together AI.
Check model cards on the Hub: you can now, in 1 click, use inference from various providers (cf video demo)
Their inference can also be used through our Inference API client. There, you can use either your custom provider key, or your HF token, then billing will be handled directly on your HF account, as a way to centralize all expenses.
๐ธ Also, PRO users get 2$ inference credits per month!
Read more in the announcement ๐ https://huggingface.co/blog/inference-providers
- 1 reply
Post
8095
Yes, DeepSeek R1's release is impressive. But the real story is what happened in just 7 days after:
- Original release: 8 models, 540K downloads. Just the beginning...
- The community turned those open-weight models into +550 NEW models on Hugging Face. Total downloads? 2.5Mโnearly 5X the originals.
The reason? DeepSeek models are open-weight, letting anyone build on top of them. Interesting to note that the community focused on quantized versions for better efficiency & accessibility. They want models that use less memory, run faster, and are more energy-efficient.
When you empower builders, innovation explodes. For everyone. ๐
The most popular community model? @bartowski 's DeepSeek-R1-Distill-Qwen-32B-GGUF version โ 1M downloads alone.
- Original release: 8 models, 540K downloads. Just the beginning...
- The community turned those open-weight models into +550 NEW models on Hugging Face. Total downloads? 2.5Mโnearly 5X the originals.
The reason? DeepSeek models are open-weight, letting anyone build on top of them. Interesting to note that the community focused on quantized versions for better efficiency & accessibility. They want models that use less memory, run faster, and are more energy-efficient.
When you empower builders, innovation explodes. For everyone. ๐
The most popular community model? @bartowski 's DeepSeek-R1-Distill-Qwen-32B-GGUF version โ 1M downloads alone.
- 4 replies
eliebakย
updated
a
Space
11 days ago
Post
9875
We are reproducing the full DeepSeek R1 data and training pipeline so everybody can use their recipe. Instead of doing it in secret we can do it together in the open!
๐งช Step 1: replicate the R1-Distill models by distilling a high-quality reasoning corpus from DeepSeek-R1.
๐ง Step 2: replicate the pure RL pipeline that DeepSeek used to create R1-Zero. This will involve curating new, large-scale datasets for math, reasoning, and code.
๐ฅ Step 3: show we can go from base model -> SFT -> RL via multi-stage training.
Follow along: https://github.com/huggingface/open-r1
๐งช Step 1: replicate the R1-Distill models by distilling a high-quality reasoning corpus from DeepSeek-R1.
๐ง Step 2: replicate the pure RL pipeline that DeepSeek used to create R1-Zero. This will involve curating new, large-scale datasets for math, reasoning, and code.
๐ฅ Step 3: show we can go from base model -> SFT -> RL via multi-stage training.
Follow along: https://github.com/huggingface/open-r1
- 5 replies
Post
2837
Today we make the biggest release in smolagents so far: ๐๐ฒ ๐ฒ๐ป๐ฎ๐ฏ๐น๐ฒ ๐๐ถ๐๐ถ๐ผ๐ป ๐บ๐ผ๐ฑ๐ฒ๐น๐, ๐๐ต๐ถ๐ฐ๐ต ๐ฎ๐น๐น๐ผ๐๐ ๐๐ผ ๐ฏ๐๐ถ๐น๐ฑ ๐ฝ๐ผ๐๐ฒ๐ฟ๐ณ๐๐น ๐๐ฒ๐ฏ ๐ฏ๐ฟ๐ผ๐๐๐ถ๐ป๐ด ๐ฎ๐ด๐ฒ๐ป๐๐! ๐ฅณ
Our agents can now casually open up a web browser, and navigate on it by scrolling, clicking elements on the webpage, going back, just like a user would.
The demo below shows Claude-3.5-Sonnet browsing GitHub for task: "Find how many commits the author of the current top trending repo did over last year."
Hi @mlabonne !
Go try it out, it's the most cracked agentic stuff I've seen in a while ๐คฏ (well, along with OpenAI's Operator who beat us by one day)
For more detail, read our announcement blog ๐ https://huggingface.co/blog/smolagents-can-see
The code for the web browser example is here ๐ https://github.com/huggingface/smolagents/blob/main/examples/vlm_web_browser.py
Our agents can now casually open up a web browser, and navigate on it by scrolling, clicking elements on the webpage, going back, just like a user would.
The demo below shows Claude-3.5-Sonnet browsing GitHub for task: "Find how many commits the author of the current top trending repo did over last year."
Hi @mlabonne !
Go try it out, it's the most cracked agentic stuff I've seen in a while ๐คฏ (well, along with OpenAI's Operator who beat us by one day)
For more detail, read our announcement blog ๐ https://huggingface.co/blog/smolagents-can-see
The code for the web browser example is here ๐ https://github.com/huggingface/smolagents/blob/main/examples/vlm_web_browser.py
- 3 replies
Post
1508
๐๐ป๐๐ฟ๐ผ๐ฑ๐๐ฐ๐ถ๐ป๐ด ๐๐ต๐ฒ ๐๐ผ๐ฟ๐น๐ฑ'๐ ๐๐บ๐ฎ๐น๐น๐ฒ๐๐ ๐๐ถ๐๐ถ๐ผ๐ป ๐น๐ฎ๐ป๐ด๐๐ฎ๐ด๐ฒ ๐บ๐ผ๐ฑ๐ฒ๐น!
Weโre thrilled to share ๐ฆ๐บ๐ผ๐น๐ฉ๐๐ (256M & 500M)โthe smallest Visual Language Models ever built. Think: running on <1GB of GPU memoryโyou can fine-tune it on your laptop and run it on your toaster!
Why Itโs Game-Changing:
- ๐ข๐๐๐ฝ๐ฒ๐ฟ๐ณ๐ผ๐ฟ๐บ๐ ๐๐ฎ๐ฟ๐ด๐ฒ๐ฟ ๐ ๐ผ๐ฑ๐ฒ๐น๐: Even the 256M model surpasses our SOTA 80B-parameter model from just 17 months ago. Over 300x reduction!
๐ ๐ถ๐ด๐ต๐๐ ๐๐ณ๐ณ๐ถ๐ฐ๐ถ๐ฒ๐ป๐ฐ๐: The 256M version delivers 80% of our 2.2B modelโs performance, and the 500M version hits 90%
๐๐ถ๐ด๐ต๐๐ป๐ถ๐ป๐ด-๐๐ฎ๐๐ ๐ฆ๐ฒ๐ฎ๐ฟ๐ฐ๐ต: SmolVLM integrates with ColiPali for state-of-the-art retrieval speedsโon par with models 10x bigger. That means cheaper, faster indexing and real-world impact.
Whatโs New Under the Hood:
- ๐ก๐ฒ๐ ๐ฉ๐ถ๐๐ถ๐ผ๐ป ๐๐ป๐ฐ๐ผ๐ฑ๐ฒ๐ฟ: Smaller overall size (400M -> 93M), but with higher resolution.
- ๐๐ถ๐ด๐ต๐ฒ๐ฟ ๐ฃ๐ถ๐ ๐ฒ๐น๐/๐ง๐ผ๐ธ๐ฒ๐ป: 4096 vs. 1820โmore efficient image processing.
- ๐ฆ๐บ๐ฎ๐ฟ๐ ๐ง๐ผ๐ธ๐ฒ๐ป๐ถ๐๐ฎ๐๐ถ๐ผ๐ป: Faster training and a performance boost.
Check our blog: https://huggingface.co/blog/smolervlm
The models: HuggingFaceTB/smolvlm-256m-and-500m-6791fafc5bb0ab8acc960fb0
The demo: HuggingFaceTB/SmolVLM-256M-Demo
Weโre thrilled to share ๐ฆ๐บ๐ผ๐น๐ฉ๐๐ (256M & 500M)โthe smallest Visual Language Models ever built. Think: running on <1GB of GPU memoryโyou can fine-tune it on your laptop and run it on your toaster!
Why Itโs Game-Changing:
- ๐ข๐๐๐ฝ๐ฒ๐ฟ๐ณ๐ผ๐ฟ๐บ๐ ๐๐ฎ๐ฟ๐ด๐ฒ๐ฟ ๐ ๐ผ๐ฑ๐ฒ๐น๐: Even the 256M model surpasses our SOTA 80B-parameter model from just 17 months ago. Over 300x reduction!
๐ ๐ถ๐ด๐ต๐๐ ๐๐ณ๐ณ๐ถ๐ฐ๐ถ๐ฒ๐ป๐ฐ๐: The 256M version delivers 80% of our 2.2B modelโs performance, and the 500M version hits 90%
๐๐ถ๐ด๐ต๐๐ป๐ถ๐ป๐ด-๐๐ฎ๐๐ ๐ฆ๐ฒ๐ฎ๐ฟ๐ฐ๐ต: SmolVLM integrates with ColiPali for state-of-the-art retrieval speedsโon par with models 10x bigger. That means cheaper, faster indexing and real-world impact.
Whatโs New Under the Hood:
- ๐ก๐ฒ๐ ๐ฉ๐ถ๐๐ถ๐ผ๐ป ๐๐ป๐ฐ๐ผ๐ฑ๐ฒ๐ฟ: Smaller overall size (400M -> 93M), but with higher resolution.
- ๐๐ถ๐ด๐ต๐ฒ๐ฟ ๐ฃ๐ถ๐ ๐ฒ๐น๐/๐ง๐ผ๐ธ๐ฒ๐ป: 4096 vs. 1820โmore efficient image processing.
- ๐ฆ๐บ๐ฎ๐ฟ๐ ๐ง๐ผ๐ธ๐ฒ๐ป๐ถ๐๐ฎ๐๐ถ๐ผ๐ป: Faster training and a performance boost.
Check our blog: https://huggingface.co/blog/smolervlm
The models: HuggingFaceTB/smolvlm-256m-and-500m-6791fafc5bb0ab8acc960fb0
The demo: HuggingFaceTB/SmolVLM-256M-Demo
- 1 reply
Post
1740
What's at stake with Meta's decision to change its content moderation policy?
@giadap
has, by far, the most thoughtful take Iโve seen on this question. Read her op-ed: https://www.techpolicy.press/when-freedom-bites-back-meta-moderation-and-the-limits-of-tolerance/
Post
1817
Reminder: Donโt. Use. ChatGPT. As. A. Calculator. Seriously. ๐ค
Loved listening to @sasha on Hard Forkโit really made me think.
A few takeaways that hit home:
- Individual culpability only gets you so far. The real priority: demanding accountability and transparency from companies.
- Evaluate if generative AI is the right tool for certain tasks (like search) before using it.
Curious about the full conversation? https://www.nytimes.com/2025/01/17/podcasts/hardfork-tiktok-rednote-environment.html. Give it a listenโitโs worth it! ๐
Loved listening to @sasha on Hard Forkโit really made me think.
A few takeaways that hit home:
- Individual culpability only gets you so far. The real priority: demanding accountability and transparency from companies.
- Evaluate if generative AI is the right tool for certain tasks (like search) before using it.
Curious about the full conversation? https://www.nytimes.com/2025/01/17/podcasts/hardfork-tiktok-rednote-environment.html. Give it a listenโitโs worth it! ๐
- 1 reply
Post
1947
Tried my hand at simplifying the derivations of Direct Preference Optimization.
I cover how one can reformulate RLHF into DPO. The idea of implicit reward modeling is chef's kiss.
Blog: https://huggingface.co/blog/ariG23498/rlhf-to-dpo
I cover how one can reformulate RLHF into DPO. The idea of implicit reward modeling is chef's kiss.
Blog: https://huggingface.co/blog/ariG23498/rlhf-to-dpo
Post
1888
Timm โค๏ธ Transformers
Wtih the latest version of transformers you can now use any timm model with the familiar transformers API.
Blog Post: https://huggingface.co/blog/timm-transformers
Repository with examples: https://github.com/ariG23498/timm-wrapper-examples
Collection: ariG23498/timmwrapper-6777b85f1e8d085d3f1374a1
Wtih the latest version of transformers you can now use any timm model with the familiar transformers API.
Blog Post: https://huggingface.co/blog/timm-transformers
Repository with examples: https://github.com/ariG23498/timm-wrapper-examples
Collection: ariG23498/timmwrapper-6777b85f1e8d085d3f1374a1
Post
1226
๐ ๐ถ๐ป๐ถ๐ ๐ฎ๐
'๐ ๐ป๐ฒ๐ ๐ ๐ผ๐ ๐๐๐ ๐ฟ๐ฒ๐ฎ๐ฐ๐ต๐ฒ๐ ๐๐น๐ฎ๐๐ฑ๐ฒ-๐ฆ๐ผ๐ป๐ป๐ฒ๐ ๐น๐ฒ๐๐ฒ๐น ๐๐ถ๐๐ต ๐ฐ๐ ๐๐ผ๐ธ๐ฒ๐ป๐ ๐ฐ๐ผ๐ป๐๐ฒ๐
๐ ๐น๐ฒ๐ป๐ด๐๐ต ๐ฅ
This work from Chinese startup @MiniMax-AI introduces a novel architecture that achieves state-of-the-art performance while handling context windows up to 4 million tokens - roughly 20x longer than current models. The key was combining lightning attention, mixture of experts (MoE), and a careful hybrid approach.
๐๐ฒ๐ ๐ถ๐ป๐๐ถ๐ด๐ต๐๐:
๐๏ธ MoE with novel hybrid attention:
โฃ Mixture of Experts with 456B total parameters (45.9B activated per token)
โฃ Combines Lightning attention (linear complexity) for most layers and traditional softmax attention every 8 layers
๐ Outperforms leading models across benchmarks while offering vastly longer context:
โฃ Competitive with GPT-4/Claude-3.5-Sonnet on most tasks
โฃ Can efficiently handle 4M token contexts (vs 256K for most other LLMs)
๐ฌ Technical innovations enable efficient scaling:
โฃ Novel expert parallel and tensor parallel strategies cut communication overhead in half
โฃ Improved linear attention sequence parallelism, multi-level padding and other optimizations achieve 75% GPU utilization (that's really high, generally utilization is around 50%)
๐ฏ Thorough training strategy:
โฃ Careful data curation and quality control by using a smaller preliminary version of their LLM as a judge!
Overall, not only is the model impressive, but the technical paper is also really interesting! ๐
It has lots of insights including a great comparison showing how a 2B MoE (24B total) far outperforms a 7B model for the same amount of FLOPs.
Read it in full here ๐ MiniMax-01: Scaling Foundation Models with Lightning Attention (2501.08313)
Model here, allows commercial use <100M monthly users ๐ MiniMaxAI/MiniMax-Text-01
This work from Chinese startup @MiniMax-AI introduces a novel architecture that achieves state-of-the-art performance while handling context windows up to 4 million tokens - roughly 20x longer than current models. The key was combining lightning attention, mixture of experts (MoE), and a careful hybrid approach.
๐๐ฒ๐ ๐ถ๐ป๐๐ถ๐ด๐ต๐๐:
๐๏ธ MoE with novel hybrid attention:
โฃ Mixture of Experts with 456B total parameters (45.9B activated per token)
โฃ Combines Lightning attention (linear complexity) for most layers and traditional softmax attention every 8 layers
๐ Outperforms leading models across benchmarks while offering vastly longer context:
โฃ Competitive with GPT-4/Claude-3.5-Sonnet on most tasks
โฃ Can efficiently handle 4M token contexts (vs 256K for most other LLMs)
๐ฌ Technical innovations enable efficient scaling:
โฃ Novel expert parallel and tensor parallel strategies cut communication overhead in half
โฃ Improved linear attention sequence parallelism, multi-level padding and other optimizations achieve 75% GPU utilization (that's really high, generally utilization is around 50%)
๐ฏ Thorough training strategy:
โฃ Careful data curation and quality control by using a smaller preliminary version of their LLM as a judge!
Overall, not only is the model impressive, but the technical paper is also really interesting! ๐
It has lots of insights including a great comparison showing how a 2B MoE (24B total) far outperforms a 7B model for the same amount of FLOPs.
Read it in full here ๐ MiniMax-01: Scaling Foundation Models with Lightning Attention (2501.08313)
Model here, allows commercial use <100M monthly users ๐ MiniMaxAI/MiniMax-Text-01
lvwerraย
authored
a
paper
21 days ago
guipenedoย
authored
a
paper
21 days ago