Empty model card: no benchmarks, no param count, same-day upload. A shame, because Apache 2.0 reasoning weights for Ascend NPUs are scarce — most open releases assume CUDA. Hard to see infra leads picking this over a documented DeepSeek distill until someone posts Ascend evals.
Unlocking Reasoning on Huawei’s Ascend with the New Apache 2.0 Asterope-21-OpenR1
The open-source reasoning landscape is expanding, and for developers working with Huawei's Ascend ecosystem, there is exciting news. A new model, Asterope-21-OpenR1, has appeared on Modelers.cn,…
aichina.news
November 7, 2026 at 6:38 PM
Empty model card: no benchmarks, no param count, same-day upload. A shame, because Apache 2.0 reasoning weights for Ascend NPUs are scarce — most open releases assume CUDA. Hard to see infra leads picking this over a documented DeepSeek distill until someone posts Ascend evals.
No benchmarks, no confirmed Ascend support — the model card is a stub. The draw: multi-sample ground-truth filtering plus LLM verification, under Apache-2.0. Hard to see this beating answer-matching until evals land, though OpenR1 distillers who've run both may disagree.
Huawei's Ascend Ecosystem Embraces OpenR1: A Cleaner Qwen2.5 Reasoning Model
The open-source community is rapidly converging on the DeepSeek-R1 style of reasoning, and the latest contribution from Modelers.cn offers a distinct take on how to achieve it. This new model,…
aichina.news
October 31, 2026 at 7:02 PM
No benchmarks, no confirmed Ascend support — the model card is a stub. The draw: multi-sample ground-truth filtering plus LLM verification, under Apache-2.0. Hard to see this beating answer-matching until evals land, though OpenR1 distillers who've run both may disagree.
Apache-2.0 and rejection-sampled ground truth make the data recipe worth studying, but with no benchmarks and a model card this thin, it is a methodology demo rather than a baseline. Anyone who has actually run it on Ascend — I would want to hear whether it holds up.
OpenR1 Gains Momentum: New Qwen2.5 Reasoning Model on Modelers.cn
The open-source race to replicate advanced reasoning capabilities is accelerating, and Modelers.cn is once again providing a critical piece of the puzzle. We are keen to highlight a new experimental…
aichina.news
October 31, 2026 at 7:44 PM
Apache-2.0 and rejection-sampled ground truth make the data recipe worth studying, but with no benchmarks and a model card this thin, it is a methodology demo rather than a baseline. Anyone who has actually run it on Ascend — I would want to hear whether it holds up.
Empty model card, no benchmarks, no parameter count — no evidence at all, just an Apache-2.0 Qwen2.5 fine-tune trained on majority-consensus data verified by OpenR1-Math-Verify, with Ascend support. Hard to see it beating Qwen2.5-Math until someone publishes a MATH score.
Unlocking Reliable Reasoning: A New Qwen2.5 Maths Model from Modelers.cn
There is a quiet revolution happening in open-source reasoning. While much of the AI community focuses on general text generation, fewer projects are tackling the arduous task of mathematical logic.…
aichina.news
October 31, 2026 at 5:26 PM
Empty model card, no benchmarks, no parameter count — no evidence at all, just an Apache-2.0 Qwen2.5 fine-tune trained on majority-consensus data verified by OpenR1-Math-Verify, with Ascend support. Hard to see it beating Qwen2.5-Math until someone publishes a MATH score.
Introducing OpenR1-Math-220k!
huggingface.co/datasets/ope...
The community has been busy distilling DeepSeek-R1 from inference providers, but we decided to have a go at doing it ourselves from scratch 💪
More details in 🧵
huggingface.co/datasets/ope...
The community has been busy distilling DeepSeek-R1 from inference providers, but we decided to have a go at doing it ourselves from scratch 💪
More details in 🧵
open-r1/OpenR1-Math-220k · Datasets at Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
huggingface.co
February 10, 2025 at 6:09 PM
Introducing OpenR1-Math-220k!
huggingface.co/datasets/ope...
The community has been busy distilling DeepSeek-R1 from inference providers, but we decided to have a go at doing it ourselves from scratch 💪
More details in 🧵
huggingface.co/datasets/ope...
The community has been busy distilling DeepSeek-R1 from inference providers, but we decided to have a go at doing it ourselves from scratch 💪
More details in 🧵
If Gemini got all the numbers right, it was like this for the 7/8B models:
"The model saw ~87% of the training data exactly once. It did not even finish one full epoch."
Training set being openR1 Math-220k.
"The model saw ~87% of the training data exactly once. It did not even finish one full epoch."
Training set being openR1 Math-220k.
January 5, 2026 at 11:37 PM
If Gemini got all the numbers right, it was like this for the 7/8B models:
"The model saw ~87% of the training data exactly once. It did not even finish one full epoch."
Training set being openR1 Math-220k.
"The model saw ~87% of the training data exactly once. It did not even finish one full epoch."
Training set being openR1 Math-220k.
We also want to express our gratitude to the broader open-source community. This research was made possible by leveraging numerous publicly available resources from DeepScaleR, STILL, OpenThoughts @bespokelabs.bsky.social , OpenR1 @hf.co , LIMR, and OpenRS.
April 23, 2025 at 5:10 PM
We also want to express our gratitude to the broader open-source community. This research was made possible by leveraging numerous publicly available resources from DeepScaleR, STILL, OpenThoughts @bespokelabs.bsky.social , OpenR1 @hf.co , LIMR, and OpenRS.
Open R1 Project Update 2: Introducing OpenR1-Math-220k and Key Advancements in Mathematical Reasoning Datasets
2025-02-10 The Open R1 project continues its mission to fill in the gaps left by DeepSeek R1, with a primary focus on reconstructing its training pipeline and synthetic data. In this…
2025-02-10 The Open R1 project continues its mission to fill in the gaps left by DeepSeek R1, with a primary focus on reconstructing its training pipeline and synthetic data. In this…
Open R1 Project Update 2: Introducing OpenR1-Math-220k and Key Advancements in Mathematical Reasoning Datasets
2025-02-10 The Open R1 project continues its mission to fill in the gaps left by DeepSeek R1, with a primary focus on reconstructing its training pipeline and synthetic data. In this second update, we introduce the OpenR1-Math-220k dataset, a substantial resource aimed at improving mathematical reasoning in language models. This dataset, generated with cutting-edge technology, marks a significant step in advancing the capabilities of AI in solving complex math problems.
undercodenews.com
February 10, 2025 at 5:30 PM
Open R1 Project Update 2: Introducing OpenR1-Math-220k and Key Advancements in Mathematical Reasoning Datasets
2025-02-10 The Open R1 project continues its mission to fill in the gaps left by DeepSeek R1, with a primary focus on reconstructing its training pipeline and synthetic data. In this…
2025-02-10 The Open R1 project continues its mission to fill in the gaps left by DeepSeek R1, with a primary focus on reconstructing its training pipeline and synthetic data. In this…
低コストで超高性能な「DeepSeek-R1」に似たAIモデルを誰でも開発できるようにオープンでない部分を補完するプロジェクト「Open-R1」をHugging Faceが始動
#DeepSeekR1 #OpenR1 #OpenAIo1 #HuggingFace #ITニュース
#DeepSeekR1 #OpenR1 #OpenAIo1 #HuggingFace #ITニュース
ITちゃんねる
低コストで超高性能な「DeepSeek-R1」に似たAIモデルを誰でも開発できるようにオープンでない部分を補完するプロジェクト「Open-R1」をHugging Faceが始動
#DeepSeekR1 #OpenR1 #OpenAIo1 #HuggingFace #ITニュース
it.f-frontier.com
January 29, 2025 at 4:25 AM
低コストで超高性能な「DeepSeek-R1」に似たAIモデルを誰でも開発できるようにオープンでない部分を補完するプロジェクト「Open-R1」をHugging Faceが始動
#DeepSeekR1 #OpenR1 #OpenAIo1 #HuggingFace #ITニュース
#DeepSeekR1 #OpenR1 #OpenAIo1 #HuggingFace #ITニュース
It's close but we still have to account that the 14B distillations are not the same as the 671B model; Tulu3 on the other hand is a 405B outperforming DeepseekV3 on which R1 was trained, so OpenR1 can make a lot of inroads. I do think there are still many untapped concepts for increased optimization
January 31, 2025 at 1:51 PM
It's close but we still have to account that the 14B distillations are not the same as the 671B model; Tulu3 on the other hand is a 405B outperforming DeepseekV3 on which R1 was trained, so OpenR1 can make a lot of inroads. I do think there are still many untapped concepts for increased optimization
OpenR1-Math-Raw
OpenR1-Math-Raw is a large-scale dataset for mathematical reasoning. It consists of 516k math problems sourced from AI-MO/NuminaMath-1.5 with 1 to 8 reasoning traces generated by DeepSeek R1.
huggingface.co/datasets/ope...
OpenR1-Math-Raw is a large-scale dataset for mathematical reasoning. It consists of 516k math problems sourced from AI-MO/NuminaMath-1.5 with 1 to 8 reasoning traces generated by DeepSeek R1.
huggingface.co/datasets/ope...
open-r1/OpenR1-Math-Raw · Datasets at Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
huggingface.co
February 13, 2025 at 7:40 AM
OpenR1-Math-Raw
OpenR1-Math-Raw is a large-scale dataset for mathematical reasoning. It consists of 516k math problems sourced from AI-MO/NuminaMath-1.5 with 1 to 8 reasoning traces generated by DeepSeek R1.
huggingface.co/datasets/ope...
OpenR1-Math-Raw is a large-scale dataset for mathematical reasoning. It consists of 516k math problems sourced from AI-MO/NuminaMath-1.5 with 1 to 8 reasoning traces generated by DeepSeek R1.
huggingface.co/datasets/ope...
🚀 Exciting times in #AI development! Showing the potential of #OpenSource! While #DeepSeek as such is open source, it is not yet possible to fully reproduce the model. Now, the #OpenR1 project on #HuggingFace is on a missi... https://byzzyy.micro.blog/2025/02/06/exciting-times-in-ai-development.html
February 6, 2025 at 9:39 PM
🚀 Exciting times in #AI development! Showing the potential of #OpenSource! While #DeepSeek as such is open source, it is not yet possible to fully reproduce the model. Now, the #OpenR1 project on #HuggingFace is on a missi... https://byzzyy.micro.blog/2025/02/06/exciting-times-in-ai-development.html
Published huggingface.co/tensopolis/q..., a fine-tune of Qwen2.5-3B model in 1 epoch of the @hf.co open-r1/OpenR1-Math-220k dataset. It trained for about 50 hours on a single A100.
#DeepSeekR1 #LLMs #AI
#DeepSeekR1 #LLMs #AI
tensopolis/qwen2.5-3b-or1-tensopolis · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
huggingface.co
February 15, 2025 at 2:33 PM
Published huggingface.co/tensopolis/q..., a fine-tune of Qwen2.5-3B model in 1 epoch of the @hf.co open-r1/OpenR1-Math-220k dataset. It trained for about 50 hours on a single A100.
#DeepSeekR1 #LLMs #AI
#DeepSeekR1 #LLMs #AI
🚀 Open-R1's latest: OpenR1-Math-220k dataset with 220k verified math problems, running on 512 H100s! The Qwen-7B model nears DeepSeek's performance, while community finds quality reasoning possible with just ~1000 samples. Check it out: https://huggingface.co/blog/open-r1/update-2
February 10, 2025 at 5:29 PM
🚀 Open-R1's latest: OpenR1-Math-220k dataset with 220k verified math problems, running on 512 H100s! The Qwen-7B model nears DeepSeek's performance, while community finds quality reasoning possible with just ~1000 samples. Check it out: https://huggingface.co/blog/open-r1/update-2
with or superior to full-data tuning and open-source baseline OpenR1-Qwen-7B across three competition-level and six comprehensive mathematical benchmarks. Further experiments highlight the scalability in varying data size, efficiency during inference, [6/7 of https://arxiv.org/abs/2505.17266v1]
May 26, 2025 at 6:09 AM
with or superior to full-data tuning and open-source baseline OpenR1-Qwen-7B across three competition-level and six comprehensive mathematical benchmarks. Further experiments highlight the scalability in varying data size, efficiency during inference, [6/7 of https://arxiv.org/abs/2505.17266v1]
length-based heuristic through a weighted scheme for ranking to prioritize high-utility examples. Empirical results on OpenR1-Math-220k demonstrate that fine-tuning LLM on only 10% of the data selected by Select2Reason achieves performance competitive [5/7 of https://arxiv.org/abs/2505.17266v1]
May 26, 2025 at 6:09 AM
length-based heuristic through a weighted scheme for ranking to prioritize high-utility examples. Empirical results on OpenR1-Math-220k demonstrate that fine-tuning LLM on only 10% of the data selected by Select2Reason achieves performance competitive [5/7 of https://arxiv.org/abs/2505.17266v1]
OpenR1-Math-220k chain-of-thought (CoT) dataset. We then apply supervised fine-tuning (SFT) and reinforcement learning fine-tuning (RLFT) to a 1.5B-parameter reasoning model, training it to learn to offload the most challenging parts of its own [5/6 of https://arxiv.org/abs/2504.16379v1]
April 24, 2025 at 5:55 AM
OpenR1-Math-220k chain-of-thought (CoT) dataset. We then apply supervised fine-tuning (SFT) and reinforcement learning fine-tuning (RLFT) to a 1.5B-parameter reasoning model, training it to learn to offload the most challenging parts of its own [5/6 of https://arxiv.org/abs/2504.16379v1]
approaches. Furthermore, we develop a methodology for reconstructing long-context reasoning datasets into our iterative format, transforming OpenR1-Math into 333K training instances. Experiments across multiple model architectures demonstrate that our [5/7 of https://arxiv.org/abs/2503.06692v1]
March 11, 2025 at 7:08 AM
approaches. Furthermore, we develop a methodology for reconstructing long-context reasoning datasets into our iterative format, transforming OpenR1-Math into 333K training instances. Experiments across multiple model architectures demonstrate that our [5/7 of https://arxiv.org/abs/2503.06692v1]
The Hugging Face team is reconstructing Deepseek-R1 to create Open-R1, a 100% open source genAI model. This alternative will offer a fully open and unrestricted solution for commercial use. Stay tuned for updates!
#OpenSource #GenAI #DeepseekR1 #OpenR1
#OpenSource #GenAI #DeepseekR1 #OpenR1
February 1, 2025 at 10:20 PM
The Hugging Face team is reconstructing Deepseek-R1 to create Open-R1, a 100% open source genAI model. This alternative will offer a fully open and unrestricted solution for commercial use. Stay tuned for updates!
#OpenSource #GenAI #DeepseekR1 #OpenR1
#OpenSource #GenAI #DeepseekR1 #OpenR1