Skip to main content

JetBrains' New AI Model Mellum2.1 Doubles Throughput Under Heavy Load

JetBrains' New AI Model Mellum2.1 Doubles Throughput Under Heavy Load

JetBrains, the company behind the tools you probably use every day, has officially released Mellum2.1—a programming AI model that's all about making agent-based coding smarter and faster. And yes, it's still open-source, which means enterprises and solo developers can deploy it privately without breaking the bank.

Reinforcement Learning: The Secret Sauce

So what's new? For starters, Mellum2.1 takes reinforcement learning from a short-term experiment and turns it into the main training engine. The model has been through the wringer—specialized training across millions of sandbox environments for software engineering and algorithms. The result? It can now self-retrieve and self-verify, pinpointing why a test failed and drafting a fix without hand-holding.

Speed That Actually Matters

Here's where it gets interesting. Thanks to multi-token prediction, Mellum2.1 responds about 1.6 times faster in single-request scenarios. But the real kicker is under heavy load: official tests show its token throughput is nearly twice that of Qwen3.5-9B, a popular open-source model. That's not just a spec sheet win—it translates to serious cost savings when you're deploying code at scale.

Why You Should Care

If you're a developer, this means less waiting and more doing. If you're running a team, it means your infrastructure budget goes further. And if you're just AI-curious, it's another sign that open-source models are catching up to—and sometimes beating—the big closed-source players.

Key Points

  • Open-source and private-deployable: No vendor lock-in, no massive bills.
  • Reinforcement learning at scale: Trained in millions of sandboxes for real-world coding tasks.
  • 1.6x faster responses: Multi-token prediction speeds up single requests.
  • 2x throughput under load: Outperforms Qwen3.5-9B in high-demand scenarios.
  • Self-verification: Can identify failed tests and suggest repairs autonomously.

Image

JetBrains isn't just throwing another model into the ring—they're addressing the messy reality of agent coding. And with Mellum2.1, they're making a strong case that open-source can be both powerful and practical.