JetBrains releases an open-source coding model Mellum2.1, with some self-reported results exceeding those of Qwen3.5-9B。
Article
JetBrains open-sources Mellum2.1：LiveCodeBench Tests surpass Qwen3.5-9B Dongcha Beating AI news flash: JetBrains releases an open-source coding model Mellum2.1. Compared with this year's June release,Mellum2 the new version still has a total of 12000000000, activating only 2500000000 with each generation, but it is better at finding code issues, modifying files, and checking the results.
JetBrains did not change the model architecture, relying mainly on reinforcement learning to improve its capabilities. The team ran millions of sandbox tasks across thousands of environments. During training, the model called terminal and file-editing tools in real code repositories, receiving rewards for successfully passing tests.
JetBrains' self-reported results show that SWE-bench Verified(the real-software-issue fixing benchmark) rose from 2%for the previous version to 47%, slightly below Qwen3.5-9B's 50%. However, in the LiveCodeBench v6 coding benchmark,Mellum2.1 scored 82%, exceeding Qwen's 75.4%. JetBrains also said that under heavy load on a single H200, the new model's output throughput approaches twice that of Qwen3.5-9B.
The model weights and the GGUF quantized version are available on Hugging Face under the Apache 2.0 license, allowing developers to deploy them locally.
Original link https://m.theblockbeats.info/flash/370983
Key points
01
JetBrains releases an open-source coding model Mellum2.1。
02
Mellum2.1 The total number of parameters is 12000000000, with 2500000000 parameters activated for each generation.
03
JetBrains said that Mellum2.1 it is better at finding code issues, modifying files, and checking the results.
04
JetBrains primarily improved the model's capabilities through reinforcement learning, running millions of sandbox tasks across thousands of environments.
05
JetBrains' self-reported results show that Mellum2.1 its success rate on SWE-bench Verified rose from 2%to 47%。
06
JetBrains' self-reported results show that Mellum2.1 on LiveCodeBench v6 it scored 82%，Qwen3.5-9B score 75.4%。
07
JetBrains said that Mellum2.1 on a single H200 under heavy load, the output throughput approaches twice that Qwen3.5-9B of the other model.
08
The model weights and the GGUF quantized version are available on Hugging Face under the Apache 2.0 license, allowing developers to deploy them locally.
AI-assisted interpretation
The following is analysis, separate from reported facts. Verify important claims independently.
Mellum2.1 is an open-source coding model from JetBrains. It was trained through reinforcement learning to use terminal and file-editing tools in real code repositories, receiving rewards based on whether tests passed. The test results in the material show that it performs better than LiveCodeBench v6 on Qwen3.5-9B, but is slightly below SWE-bench Verified on Qwen3.5-9B。
Why it matters to readers
Developers can obtain the model weights and the GGUF quantized version and try deploying them locally; its code-testing and execution-efficiency performance can serve as a reference when selecting a model.
Beginners can understand it as an open-source model that can find, modify, and test code, but the performance figures in the material mainly come from JetBrains' self-reported results.
Risks and unknowns
The performance data come from JetBrains' self-reported results; the material provides no independent test results.
The material provides no official announcement, model card, or specific Hugging Face page information.
The actual performance of local deployments and throughput on different hardware remain unclear.
Related Developments
Loading event timeline…
Related concepts
SWE-bench Verified
This term is not in the glossary yet. Browse related concepts in the glossary.
