Home
Baidu Launches Ernie 5.1, Slashing Pre-training Costs by 94% and Ranking Among Top Four Globally
Baidu officially launched its next-generation language model, Ernie 5.1, on May 11, 2026. Building upon the Ernie 5.0 pre-training foundation released earlier in January—a model with 2.4 trillion parameters—this new version was developed using an innovative "one-shot elastic training framework." This approach enabled Baidu to optimize multiple model sizes in a single training cycle, slashing Ernie 5.1's pre-training costs to just 6% of comparable models.

As of May 9, Ernie 5.1 held the fourth position globally and ranked first among Chinese models on the Arena Search leaderboard with a score of 1223 points, showcasing its exceptional efficiency and well-balanced performance.
Architecturally, Ernie 5.1 features a sub-model design with tunable depth, width, and number of active experts. Its total parameter count is only one-third of its predecessor, and the number of effective parameters activated per query has been roughly halved. To address the "see-saw effect" common in multi-skill training, Baidu implemented a four-phase post-training process. This method utilizes parallel specialized training code, inference, and proxy expert models, combined with strategy distillation and reinforcement learning, effectively resolving the industry-wide challenge of balancing programming capability with creative reasoning. Furthermore, a revamped reinforcement learning infrastructure decouples model updates, response generation, and evaluation. When paired with a standardized low-precision computation library, this significantly enhances the stability of large-scale training.

Related article
DeepMind CEO Hassabis: I sleep six hours a day, usually feel energetic around 1 a.m.
Fortune recently featured an interview with Demis Hassabis, CEO of Google DeepMind, revealing his unconventional approach to rest and productivity. Hassabis disclosed that he sleeps very little, structuring his waking hours into two distinct work blo
OpenAI, Anthropic Vie for Market Share Despite Revenue Shortfalls
Despite recent reports suggesting OpenAI missed revenue targets, creating pressure on tech stocks this Tuesday, private AI lab investors remain resilient. Seasoned backers have confirmed they will not reduce investment despite negative media coverage
California AV Compliance: A New Era of Tickets, Geofences, and 1M Miles
Guident operates an AuveTech shuttle in South Florida, managing a four-mile route in West Palm Beach and a one-mile route in Boca Raton using its remote monitoring technology. | Credit: GuidentCalifornia is redefining the regulatory landscape for dri
Related Special Topic Recommendations
Comments (0)
0/500
Baidu officially launched its next-generation language model, Ernie 5.1, on May 11, 2026. Building upon the Ernie 5.0 pre-training foundation released earlier in January—a model with 2.4 trillion parameters—this new version was developed using an innovative "one-shot elastic training framework." This approach enabled Baidu to optimize multiple model sizes in a single training cycle, slashing Ernie 5.1's pre-training costs to just 6% of comparable models.

As of May 9, Ernie 5.1 held the fourth position globally and ranked first among Chinese models on the Arena Search leaderboard with a score of 1223 points, showcasing its exceptional efficiency and well-balanced performance.
Architecturally, Ernie 5.1 features a sub-model design with tunable depth, width, and number of active experts. Its total parameter count is only one-third of its predecessor, and the number of effective parameters activated per query has been roughly halved. To address the "see-saw effect" common in multi-skill training, Baidu implemented a four-phase post-training process. This method utilizes parallel specialized training code, inference, and proxy expert models, combined with strategy distillation and reinforcement learning, effectively resolving the industry-wide challenge of balancing programming capability with creative reasoning. Furthermore, a revamped reinforcement learning infrastructure decouples model updates, response generation, and evaluation. When paired with a standardized low-precision computation library, this significantly enhances the stability of large-scale training.

DeepMind CEO Hassabis: I sleep six hours a day, usually feel energetic around 1 a.m.
Fortune recently featured an interview with Demis Hassabis, CEO of Google DeepMind, revealing his unconventional approach to rest and productivity. Hassabis disclosed that he sleeps very little, structuring his waking hours into two distinct work blo
OpenAI, Anthropic Vie for Market Share Despite Revenue Shortfalls
Despite recent reports suggesting OpenAI missed revenue targets, creating pressure on tech stocks this Tuesday, private AI lab investors remain resilient. Seasoned backers have confirmed they will not reduce investment despite negative media coverage











