Reportedly: JetBrains releases Mellum2.1 coding model with near 2x Qwen3.5-9B throughput under load
Reported by IT Home · not yet confirmed by the company or a second independent outlet. We update this page when it is.
JetBrains released Mellum2.1, a 12B MoE coding model with 2.5B active parameters under Apache 2.0, on October 8, 2026. It reports near 2x the high-load inference throughput of Qwen3.5-9B and 1.6x faster single-request speed via multi-token prediction.
China context
- Outside China
- Not stated in the sources yet
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Builders outside China can benchmark their coding models against Mellum2.1's reported throughput and agentic coding capabilities, but should verify independently.
- For investors
- Investors can monitor whether JetBrains' open-weight release pressures commercial coding model pricing or adoption.
Translated from Chinese. Quotes and facts link to the original sources.
JetBrains announced Mellum2.1 on October 8, 2026, upgrading the reinforcement learning stage of its Mellum2 model to improve agentic coding. The model keeps the 12B mixture-of-experts architecture with 2.5B active parameters and Apache 2.0 license. Training used millions of sandboxes across thousands of environments and filtered open datasets. JetBrains reports the largest gains in agentic coding: exploring codebases, editing files, identifying root causes of failing tests, drafting fixes, and verifying results. Multi-token prediction increases single-request speed by about 1.6x, while high-load inference throughput is nearly twice that of Qwen3.5-9B.
Mellum2.1 shifts reinforcement learning from a short post-training step to the main training phase, with data added for math, algorithmic competitions, science, tool use, and software engineering. The model uses multi-token prediction to improve response speed without changing architecture. Reported throughput under high load is nearly 2x Qwen3.5-9B, but this is company-reported and not independently verified.
Developers using JetBrains AI coding tools gain a faster, more agentic model at no additional license cost, while open-weight competitors face a new Apache 2.0 baseline for coding throughput. The next verifiable signal is whether independent benchmarks or third-party evaluations confirm the reported throughput and agentic coding gains.
JetBrains strengthens its AI coding assistant with a faster, more capable model under a permissive license, potentially reducing inference costs and improving developer productivity for its user base.
The next observable signal is independent evaluation of Mellum2.1's agentic coding and throughput claims, or adoption metrics from JetBrains AI users. No evidence yet shows availability outside JetBrains products or impact on non-JetBrains developers.