JetBrains has released Mellum2 as an open-weight coding model, making a clear bet that faster, IDE-native AI will matter as much as larger general models.
JetBrains is no longer just adding AI features around its developer tools. With Mellum2, it is putting a serious model into the open and asking developers to judge whether a focused system built for software work can compete with the broader assistants now fighting for the coding workflow.
The release matters because JetBrains already owns a valuable piece of the developer day. IntelliJ IDEA, PyCharm, WebStorm and the rest of its IDE family sit where code is written, reviewed, renamed, navigated and refactored. If AI assistance is going to become a routine part of that work, JetBrains has every reason to make the model feel native to the editor rather than bolted on through a generic chat window.
According to the Mellum2 technical report published on arXiv, the model is a 12 billion parameter Mixture-of-Experts system with 2.5 billion active parameters per token, released under the Apache 2.0 license. That is the key design choice. JetBrains is not trying to win by making the largest model in the room. It is trying to make a model that can do useful software engineering work quickly enough to fit into the daily rhythm of an IDE.
Mellum2 follows the first Mellum model, a 4 billion parameter dense model built mainly for code completion. The new version is broader. JetBrains describes it as a general-purpose software engineering model that can handle code generation and editing, debugging, multi-step reasoning, tool use, function calling, agentic coding and conversational programming assistance.
That distinction is important. Code completion is still valuable, but the market has moved beyond filling in the next line. Developers now expect AI tools to make edits across files, explain failures, reason through unfamiliar code, call tools and work through longer tasks without losing context. Cursor, GitHub Copilot and API-first coding agents have reset expectations very quickly.
Mellum2 is built around efficiency. Its architecture uses 64 experts with 8 active at a time, grouped-query attention, sliding window attention on most layers and a multi-token prediction head designed to help speculative decoding. The model also supports a 128K context window after long-context extension, which gives it room to work with more of a project than a narrow autocomplete engine could manage.
For a developer, those details only matter if they change the feel of the tool. A coding assistant that takes too long breaks concentration. A model that needs expensive remote inference for every interaction creates cost and privacy concerns. A smaller active-parameter footprint gives JetBrains a path toward faster responses and more flexible deployment, including local or controlled environments where companies are cautious about sending code to outside services.
Open weights are also a trust strategy
Open-sourcing Mellum2 is not charity. It is a strategic move in a market where developers are skeptical, technical and difficult to impress with vague AI promises. Open weights let researchers, advanced users and enterprise teams inspect, benchmark and adapt the model in ways that a closed assistant cannot support.
JetBrains has released several Mellum2 checkpoints on Hugging Face, including base, instruct and thinking variants. The thinking version is aimed at complex debugging, planning and reasoning-heavy work, while the instruct version is positioned for lower-latency direct answers. That gives the company a useful split: one model family for interactive coding assistance, another for deeper workflows where a few more seconds may be acceptable.
The business question is how JetBrains makes money from an open-weight core. The answer is likely not the model alone. The real product is the workflow around it: IDE integration, project context, refactoring tools, evaluation, enterprise controls and the paid JetBrains AI layer. That is where JetBrains can create value that a downloaded model file cannot easily reproduce.
This is also where JetBrains has a different kind of moat. GitHub Copilot benefits from GitHub distribution and Microsoft’s cloud reach. Cursor has momentum with developers who want an AI-first editor. Open model providers such as Meta, Mistral and Google are pushing strong general-purpose systems into the market. JetBrains has something more specific: decades of IDE behavior, code intelligence and language tooling built around how professional developers actually move through projects.
That does not guarantee success. Developers will not tolerate weaker results just because a model is nicely integrated. If Mellum2 falls behind on real coding tasks, the open-source release will mostly serve researchers and local-model enthusiasts. But if it is fast enough, private enough and good enough inside the editor, it gives JetBrains a practical answer to the growing fear that standalone coding agents will make traditional IDEs less central.
The next test is adoption. Watch whether Mellum2 gains community ports, quantized builds and third-party benchmarks, but also watch how quickly JetBrains folds the model into everyday AI Assistant features. The model release is only the start. The bigger story is whether the company can turn open weights into a tighter, more trusted developer workflow.
Also read: Anthropic has started the public market test for frontier AI • AI weather startups are turning forecasting into a private market • AI is forcing economists to recheck the math beneath markets
