Ivan Bondyrev

Research Engineer @jetbrains, Mellum

[Ivan Bondyrev @ivanbondyrev]

@jaswwei Nice one! What about MOPD? Do u think it is a reliable way to get more signal?

[Ivan Bondyrev @ivanbondyrev]

@nv_pavlichenko @jetbrains We're gonna miss you 😭

[Nikita Pavlichenko @nv_pavlichenko]

Finally releasing the results of the arch ablations I was running with bux over the weekend https://t.co/S5VSc51uxN

[PyCharm, a JetBrains IDE @pycharm]

Meet Mellum2.1 – JetBrains' open-source 12B MoE model built for coding agents and local execution.

🚀 140+ T/S in local testing scenarios
🇪🇺 Ethically trained European open model
💻 Efficient worker model inside IDE agent workflows

Try it locally: https://t.co/7JuOGEDAHC

[Ivan Bondyrev @ivanbondyrev]

@nv_pavlichenko Proud of the team, and lucky with the lead who kept us pointed in the right direction.

[Ivan Bondyrev @ivanbondyrev]

The most satisfying part was watching our RL infrastructure and training protocol mature over the summer. I can only imagine what we can build on this foundation next.

[Ivan Bondyrev @ivanbondyrev]

Mellum2.1 is out.

A major update to the Mellum2 we open-sourced in June. The architecture is unchanged, 12B mixture-of-experts with 2.5B active parameters under Apache 2.0. What is new is the scale of reinforcement learning in post-training: millions of sandboxed runs inside real repositories, aimed at agentic coding.

The result is a small model that works as a coding agent. On SWE-bench Verified we achieved 47%, with gains across coding, math, tool use, and general knowledge as well.

It is also the fastest model in its class. Under load it serves about twice the tokens of Qwen3.5-9B on a single H200.

Built for sub-agents, coding agents and private self-hosted deployment. Weights are on Hugging Face: both HF and GGUF formats. MTP support is on its way.

Blog: https://t.co/VZn3tmlQ2e
🤗 collection: https://t.co/EvtnRyHh2d

Proud of this one. Try it and tell us what breaks.

[Ivan Bondyrev @ivanbondyrev]

That infrastructure unlocked full-scale training: millions of sandboxed runs feeding asynchronous RL with a CISPO loss, which we found gives noticeably more stable training than GRPO at this scale.

[Ivan Bondyrev @ivanbondyrev]

The biggest challenge this time was scaling RL itself. Millions of sandboxes need constant attention: provisioning, health, cleanup, and getting results back to the trainer in time. We built the infrastructure for hosting and managing them as a separate system, fully decoupled from training, which let us scale it horizontally by simply adding machines, without touching the training job.

[Viktor Tiulpin 🇺🇦 @tiulpin]

please ddos junie lite (legally, by using it, it's free)

i'm on call today and haven't written a postmortem in weeks. getting rusty