Mellum lands: JetBrains' first open AI coding model
Mellum has taken the wraps off as JetBrains' first "open" AI model for coding, placing a 4-billion-parameter code-completion system on Hugging Face under an Apache 2.0 license. The move marks the toolmaker's first public step away from proprietary large language models and into transparent, community‑driven distribution. As JetBrains' Jodie Burchell put it, "Mellum is the first in a growing family of focal models where each is designed around a specific developer task." The model is not a general‑purpose chatbot; it is trained from scratch to do one job well: code completion.
A 4‑billion‑parameter code model
Mellum weighs in at 4 billion parameters, a size that places it squarely in the territory of serious code‑generation systems. According to the TechCrunch report, the parameter count corresponds to a model's problem‑solving skills, while the 4 trillion tokens it was trained on equate to roughly 120 million lines of code — enough context to capture many patterns found in real‑world software projects. The training data mixed permissively licensed code from GitHub with English‑language Wikipedia articles, giving the model both programming syntax and the general language fluency that often helps with code‑related queries.
Trained on 4 trillion tokens of code and text
The scale of the training run is hard to overstate. JetBrains used a cluster of 256 H200 Nvidia GPUs for about 20 days to finish the training. That compute budget bought a model that can ingest and reason over a massive corpus of public code, something the company says is necessary if the model is to give useful completions in a variety of languages and frameworks. The TechCrunch article notes that "a million tokens is equivalent to ~30,000 lines of code," so 4 trillion tokens represents a dataset an order of magnitude larger than what many smaller code‑model papers publish.
Apache 2.0 license, freely available
One of the most striking aspects of the release is the Apache 2.0 license. Unlike many AI models that ship with restrictive terms or hidden commercial clauses, Mellum's license permits commercial use, modification, and distribution without the kind of legal friction that often stymies open‑source AI projects. The TechCrunch piece emphasizes that this was a deliberate choice: JetBrains wants researchers, educators, and independent developers to tinker with the weights, fine‑tune on private codebases, and contribute back improvements. The JetBrains Guide page confirms that the model was "open‑sourced on Hugging Face to invite transparency, collaboration, and contribution."
Fine‑tuning required before production use
Mellum cannot be dropped into a production IDE out of the box. The base model needs a fine‑tuning step before it can give reliable completions for a specific codebase or language. JetBrains itself provided a few Python‑fine‑tuned variants, but the company is explicit: those are "meant for estimation about potential capabilities," not for shipping real code. The TechCrunch article quotes the JetBrains blog post saying the base model "can't be used out of the box; it has to be fine‑tuned first." This is a key practical detail for anyone hoping to experiment with Mellum immediately.
Bias and security considerations
Because Mellum was trained on public codebases, it inherits the patterns — good and bad — present in those repositories. The TechCrunch report flags two specific concerns. First, the model "may reflect biases present in public codebases," meaning its suggestions could mirror the stylistic quirks, naming conventions, or even deprecated practices found in the training data. Second, JetBrains warns that its code suggestions "won't necessarily be secure or free of vulnerabilities." A 2023 Snyk survey cited in the same article found that more than 50 % of organizations encounter security issues with AI‑produced code "sometimes or frequently." Those caveats are not meant to discourage experimentation but to set realistic expectations.
JetBrains' focal model strategy
Mellum is the opening salvo in what JetBrains calls its "focal model" family. The philosophy is straightforward: not every AI task requires a general‑purpose large model. Some jobs — code completion, diff prediction, refactoring assistance — are better served by a model that is small, fast, and finely focused. The JetBrains Guide page describes Mellum as "a purpose-built language model trained from scratch to do one job well: code completion. It's fast, lightweight, and focused, prioritizing depth over breadth." The guide also teases other focal models in the works, from diff‑prediction to other developer‑task–specific capabilities. The overarching angle, as the TechCrunch article puts it, is that JetBrains "is not chasing generality — we're building focus."
Try Mellum in JetBrains IDEs
For developers who want to kick the tires without setting up a local fine‑tuning pipeline, JetBrains has integrated a version of Mellum directly into its IDEs. The JetBrains AI Assistant, available in PyCharm and the broader IDE suite, already runs a proprietary version of the model. The Guide page includes a "Try it in your JetBrains IDE" call‑to‑action, pointing to the built‑in AI features that already use Mellum under the hood. This means many JetBrains users are already interacting with the model, even if they haven't noticed the name.
Explore on Hugging Face
The most immediate way to get hands‑on with Mellum is to visit its Hugging Face page, where the model weights, tokenizer, and example notebooks are posted. The JetBrains Guide mentions "Explore it on Hugging Face" as a direct link, and the TechCrunch article notes that the model went live on the platform on Wednesday, April 30, 2025. From there, anyone can download the base checkpoint, run the provided fine‑tuning scripts, or simply probe the model's capabilities with a few sample prompts. The openness of the release — Apache 2.0, full model cards, community‑friendly documentation — is precisely the setup JetBrains says it hopes will spark "even one meaningful experiment, contribution, or collaboration."
If you're reading this and wondering whether Mellum can replace your current code‑completion tool: it probably can't yet. But as the first open‑licensed model from a major IDE vendor, it's a signal that the era of completely closed AI assistants is ending, and the next wave will be shaped by whoever invites the community in first.
