JetBrains recently announced that Mellum2, its next-generation machine learning model for software engineering systems, has been officially open sourced. This is just over a year after the company open sourced its first Mellum model in 2025. At that time, Mellum was a small model focused on code completion, with a parameter size of approximately 4 billion. In comparison, the total number of parameters of Mellum2 has expanded to 12 billion.However, JetBrains emphasizes that by adopting a sparse Mixture-of-Experts (MoE) architecture, each token actually participates in the calculation of only about 2.5 billion active parameters, so that it can still maintain fast inference speed on standard hardware.
Unlike "focused" Mellum, which only performs code completion tasks in the editor, Mellum2 is positioned as a complete intelligent coding assistant. It can generate and edit code, call external tools, execute multi-step "Agentic" workflows, support long conversations, and have explicit reasoning capabilities. Developers can use its base, directive-tuned, and "Think Edition" models under the Apache 2.0 license to maintain full control over their data when building products or internal tools.
In terms of specifications, Mellum2 has significant changes in multiple key indicators compared to the first generation Mellum: the original Mellum used a dense parameter design, while Mellum2 switched to a MoE architecture; the total number of parameters increased from 4 billion to 12 billion, but the active parameters dropped from 4 billion to 2.5 billion; the context window increased significantly from 8192 tokens to 131072 tokens, making it more suitable for processing long documents and complex engineering scenarios; both generations of models use Apache 2.0 Open Source License.
JetBrains said Mellum2 is particularly suitable for routing and orchestrating complex AI workloads, building low-latency retrieval-augmented generation (RAG) pipelines, providing fast inference support for sub-agents in large systems, and deploying AI in local or private environments. In order to further improve the model's performance in professional fields such as code and mathematics, the team adopted a three-stage data "course" in the pre-training stage, gradually migrating from diverse web data to more carefully screened code and mathematics content.
In addition to Mellum2, JetBrains also announced a product line adjustment: the company will stop maintaining DataSpell, its data science IDE for Python developers. According to official information, DataSpell has been officially "retired" on May 28, and related data science functions will be integrated into the PyCharm professional version. Existing active DataSpell subscriptions will be automatically converted to PyCharm Pro on September 1, 2026.