For edge applications (it's a 1B model!), this is not enough. We need more concise traces. With a 4k token budget, the GRPO (LFM-1.3B-Math) model achieves best-in-class performance. Note that it only cost us a few points on average at the 32k token setting.
MACHINE LEARNING
-
LFM-1.3B-Distill SFT Model Outperforms Larger DeepSeek Competitor
By
–
Evals show that our SFT (LFM-1.3B-Distill) model performs slightly better with a 32k token budget. It's actually competitive with models based on DeepSeek-R1-Distill-Qwen-1.5B while being 15% smaller.
-
GRPO with Difficulty-Aware Rewards for AI Model Training
By
–
We leveraged GRPO with a difficulty-aware reward formulation to address this issue. (More information about our custom GRPO flavor in the article.) We combined it with the following data mix after filtering out samples that do not yield a solve rate of 20-80% at 4k tokens.
-
SFT Fine-tuning Strategy for 1.3B Chat Model
By
–
The first step is SFTing our general-purpose 1.3B chat model on the right mixture. We iterated over many open-source datasets to find the ideal mix. This is great to raise performance, but also makes the model extremely verbose (>10k tokens on average).
-
Small 1B Model Achieves Strong Reasoning Performance
By
–
Can small models be strong reasoners? We created a 1B reasoning model at @liquidai that is both accurate and concise We applied a combination of SFT (to raise quality) and GRPO (to control verbosity) The result is a best-in-class model without specific math pre-training
-
Andy Konwinski Launches Laude Institute With $100 Million
By
–
Databricks & Perplexity co-founder Andy Konwinski launched the Laude Institute with $100M
— Rowan Cheung (@rowancheung) 25 juin 2025
The non-profit will convene, resource, and amplify computer scientists to help them fast-track breakthroughs for real-world impactpic.twitter.com/PZSZDaLCOhDatabricks & Perplexity co-founder Andy Konwinski launched the Laude Institute with $100M The non-profit will convene, resource, and amplify computer scientists to help them fast-track breakthroughs for real-world impact
-
Google DeepMind Releases Gemini Robotics On-Device VLA Model
By
–
Google DeepMind dropped Gemini Robotics On-Device, a VLA model that can run locally on robots — without internet
— Rowan Cheung (@rowancheung) 25 juin 2025
It can handle a wide variety of complex, two-handed tasks out of the box and even learn new skills with demonstrationspic.twitter.com/FA3YdkMdGHGoogle DeepMind dropped Gemini Robotics On-Device, a VLA model that can run locally on robots — without internet It can handle a wide variety of complex, two-handed tasks out of the box and even learn new skills with demonstrations
-
BAIR researchers win Outstanding Demo Paper Award at RSS 2025
By
–
Congratulations to BAIR researchers @kevin_zakka @qiayuanliao @arthurallshire @carlo_sferrazza @KoushilSreenath @pabbeel and Google collaborators for winning the Outstanding Demo Paper Award at RSS 2025! https://
playground.mujoco.org -
BAIR Faculty Awarded Google Research Scholar Awards
By
–
Congratulations to BAIR faculty @_ahmedmalaa and @serinachang5 who have been awarded Google Research Scholar Awards under the "Health Research" Category and BAIR Alumnus Ashwin Pananjady who has been awarded the Google Research Scholar Award under the "Machine Learning"
-
Google Announces Gemini Robotics On-Device System for Autonomous Operations
By
–
Building on more than 10 years of robotics research and engineering at @GoogleDeepMind, @GoogleResearch and @GoogleAI, we're delighted to announce our Gemini Robotics On-Device system. A really capable vision-language-action model that can run entirely without network access. ⬇️ https://t.co/EBso5EBigr
— Jeff Dean (@JeffDean) 25 juin 2025Building on more than 10 years of robotics research and engineering at @GoogleDeepMind
, @GoogleResearch and @GoogleAI
, we're delighted to announce our Gemini Robotics On-Device system. A really capable vision-language-action model that can run entirely without network access.