JetBrains Releases Mellum2.1 12B MoE Coding Model
JetBrains has released Mellum2.1, an Apache 2.0 open-source 12B mixture-of-experts thinking model featuring 2.5B active parameters.
Stock photo for illustration only, not from the actual event
- JetBrains launched Mellum2.1 open-source 12B model
- Features Mixture-of-Experts with 2.5B active parameters
- SWE-bench Verified score jumped from 2.0 to 47.0
- Released under the Apache 2.0 license
JetBrains has officially introduced Mellum2.1, a cutting-edge artificial intelligence model tailored specifically for coding agents. Built on a mixture-of-experts (MoE) architecture, the model boasts a total of 12 billion parameters while utilizing 2.5 billion active parameters during operation.
A key advancement in Mellum2.1 is the application of reinforcement learning (RL) directly within real code repositories, which significantly enhances its capability to handle complex software engineering tasks.
Through rigorous evaluation, the model's performance on the SWE-bench Verified benchmark showed a dramatic improvement, surging from an initial score of 2.0 up to 47.0.
Stock photo for illustration only, not from the actual event
JetBrains' decision to leverage a mixture-of-experts architecture combined with real-repository reinforcement learning highlights the industry shift toward highly efficient and specialized coding assistants. By releasing the model under the Apache 2.0 license, developers and researchers gain transparent access to build upon and audit the technology.
Source: MarkTechPost
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment