Mellum2.1

JetBrains' 12B mixture-of-experts coding-agent model with 2.5B active parameters, a 128K context and the Apache 2.0 license.

In short. Mellum2.1 is a 12-billion-parameter model from JetBrains, the maker of IntelliJ IDEA and PyCharm, built to work inside code repositories and to act as a fast sub-agent. Only 2.5 billion parameters are active for each token, so it runs quickly, and its 4-bit files are about 7 to 8 GB. It reasons before it answers, and the license is Apache 2.0.

Summary

Mellum2.1 is the October 2026 update of Mellum2, the 12B mixture-of-experts model that JetBrains open-sourced in June 2026 as the successor to its 4B dense Mellum completion model. The architecture is unchanged: 12.15B parameters in 28 layers, 64 experts with 8 active per token (2.5B active parameters), grouped-query attention with 4 key-value heads, sliding-window attention on three of every four layers with a 1,024-token window, and a 131,072-token context. The Mellum2 technical report says the base model saw about 10.6 trillion tokens. Almost all the new work went into post-training: JetBrains says reinforcement learning became the main stage, with new tasks in math, competitive programming, science, tool use and software engineering, and millions of sandboxed runs in which the model works inside real repositories with a shell and file-editing tools and is rewarded when the tests pass. Mellum2.1 is a thinking model that writes its reasoning before the answer.

In JetBrains' table, every model was run through one pipeline in thinking mode, and Mellum2.1 scores 82.0 on LiveCodeBench v6, 83.3 on AIME 2025 and 2026 combined, 47.0 on SWE-bench Verified, 62.3 on BFCL v4 and 64.6 on GPQA Diamond. The June Mellum2 Thinking scored 2.0 on SWE-bench Verified and Gemma 4 E4B scored 23.0. Qwen3.5-9B is ahead of Mellum2.1 on SWE-bench Verified (50.0), SWE-bench Pro (38.0 against 28.0), Terminal-Bench 2.1 (21.7 against 17.4) and GPQA Diamond (77.8), and it has the lower, better HarmBench score (6.6 against 8.5). JetBrains also says Mellum2.1 serves almost twice as many tokens as Qwen3.5-9B under heavy load on one H200, and that its multi-token prediction head makes a single request about 1.6 times faster. All of these numbers are JetBrains' own.

The BF16 weights are 24.3 GB. JetBrains publishes its own GGUF files, 12.9 GB at Q8_0, 10.9 GB at Q6_K, 8.1 GB at Q4_K_M, which it recommends, and 7.0 GB at MXFP4_MOE, and the GGUF repository includes the multi-token prediction head for speculative decoding in llama.cpp. A separate repository adds the head for vLLM 0.33 or newer and Transformers 5.20 or newer. The card suggests temperature 0.6, top-p 0.95 and top-k 20, and shows vLLM commands with the qwen3 reasoning parser and the Hermes tool-call parser. JetBrains has published only a Thinking checkpoint for 2.1, so the direct-answer Instruct and the Base checkpoints are still the June Mellum2 ones. In the 30 days to 11 November 2026, Hugging Face counted about 440,000 downloads of the GGUF repository and about 12,000 of the main weights.

Good for

  • Coding agents and fast sub-agents that run on your own hardware, which JetBrains built it for
  • Laptops and single consumer GPUs, since the 8.1 GB Q4_K_M and 7.0 GB MXFP4 files fit in 16 GB of memory by our estimate
  • Hard math and reasoning problems from a small active size, where JetBrains reports 83.3 on AIME
  • Fine-tuning and research, because the June base checkpoint is published too

Watch out for

  • Every benchmark number is JetBrains' own, and Qwen3.5-9B scores higher on SWE-bench Verified, SWE-bench Pro, Terminal-Bench 2.1 and GPQA Diamond in the same table
  • It always reasons before answering, so use the June Mellum2 Instruct if you want direct answers
  • Speculative decoding needs the MTP head, which is a separate repository for vLLM and newer library versions
  • It handles text only, and its HarmBench score (8.5, lower is better) is worse than Qwen3.5-9B's

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Estimates for the Mellum2.1 Thinking GGUF MXFP4_MOE variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.

Variants Parameters BF16 FP8 / INT8 4-bit (Q4) Fits on
Mellum2.1 Thinking BF16 12.2B 2.5B active 28 GB 14 GB 8.38 GB Laptop or phone
Mellum2.1 Thinking GGUF Q4_K_M INT4 12.2B 2.5B active 28 GB 14 GB 8.91 GB Laptop or phone
Mellum2.1 Thinking GGUF MXFP4_MOE MXFP4 12.2B 2.5B active 28 GB 14 GB 7.7 GB Laptop or phone
Mellum2 Instruct (June 2026) BF16 12.2B 2.5B active 28 GB 14 GB 8.38 GB Laptop or phone
Mellum2 Base (June 2026) BF16 12.2B 2.5B active 28 GB 14 GB 8.38 GB Laptop or phone

Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.

VRAM calculator

The license in plain words

One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.

More about Apache 2.0

See also

Frequently asked questions

What is Mellum2.1?
Mellum2.1 is a 12-billion-parameter model from JetBrains, the maker of IntelliJ IDEA and PyCharm, built to work inside code repositories and to act as a fast sub-agent. Only 2.5 billion parameters are active for each token, so it runs quickly, and its 4-bit files are about 7 to 8 GB. It reasons before it answers, and the license is Apache 2.0.
Can I use Mellum2.1 commercially?
Yes. Mellum2.1 is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
Is Mellum2.1 open source?
Mellum2.1 is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
Can I run Mellum2.1 on my own computer?
Yes. The smallest variant, Mellum2.1 Thinking GGUF MXFP4_MOE, needs about 7.7 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
When was Mellum2.1 released?
Mellum2.1 was released by JetBrains on October 8, 2026.
How long a context window does Mellum2.1 have?
Mellum2.1 supports a context window of up to 128K tokens.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
Qwen3.5 Alibaba Qwen 873M – 403B 17B active 256K Apache 2.0 (Permissive) 0.602 GB Laptop or phone
Gemma 4 Google DeepMind 5.1B – 30.7B 256K Apache 2.0 (Permissive) 3.52 GB Laptop or phone
Flint Recursive 39.8B 2.5B active 256K Apache 2.0 (Permissive) 48 GB Workstation
LFM2.5-24B-A2B Liquid AI 24B 2.45B active 128K LFM Open 1.0 (Conditional) 17 GB Gaming GPU