JetBrains releases Mellum 2.1 Thinking, a small Apache-licensed coding model built to run locally

The news

JetBrains has released Mellum 2.1 Thinking, an open-weight coding model. It is a mixture-of-experts with 12 billion total parameters but only 2.5 billion active per token (64 experts, eight used at a time), a 131,072-token context window, and an Apache 2.0 licence. The weights are on Hugging Face as BF16 safetensors. It is a “thinking” model, meaning it reasons before answering, and JetBrains attributes the step up from Mellum 2 to reinforcement learning in real software-engineering environments rather than to any architectural change.

It serves today on vLLM and SGLang. JetBrains says GGUF builds for llama.cpp, Ollama and LM Studio are “coming soon”; in the meantime the model’s Hugging Face page already links around sixteen community-quantised derivatives, so in practice a GGUF build may already be a search away, even though the official ones are not out. We note that gap rather than smoothing over it.

Why it matters

The active-parameter count is the point. At 2.5 billion active, the model runs fast, and a 4-bit quantisation of the full 12 billion parameters is roughly 7GB (an approximate figure), which fits a modest GPU or a laptop with a little headroom. That makes it a genuinely local coding model, in contrast to the frontier open coders of 2026 that run to hundreds of billions of parameters and need server hardware. Apache 2.0 means no commercial conditions to read around. For someone who wants a coding assistant that runs on the machine in front of them, this is the interesting kind of release.

The benchmarks, read honestly

JetBrains reports, as its own figures on the model card (self-reported, October 2026), SWE-bench Verified 47.0, LiveCodeBench v6 82.0, AIME 25/26 83.3, GPQA Diamond 64.6, and IFEval 90.6. The SWE-bench jump is the headline: JetBrains says Mellum 2 scored 2.0 under the same harness, so 47.0 is a large generational gain.

The same card is also where the caution lives. On Terminal-Bench 2.1, an agentic terminal-coding benchmark, Mellum 2.1 scores 17.4, which JetBrains’ own table puts below a smaller rival (Qwen3.5-9B at 21.7). So this is not a clean category leader: strong on code generation and the knowledge-style benchmarks, weaker on long-horizon agentic terminal work. Treat all of these as the developer’s claims until independent evaluations appear, and remember that the agent scaffold around a model can move a coding score substantially. If you are weighing it up, our guide on how to read benchmarks and the split between code generation and agentic coding in evaluating for your use case are the relevant background.

The history

Mellum is JetBrains’ own line of coding models, built to sit behind its IDEs, and the company has released the weights openly rather than keeping them to its products. It lands in a steady run of small, permissively licensed coding models through 2026, following releases like Mistral 3 under Apache 2.0. The pattern is encouraging for local use: while the headlines go to trillion-parameter flagships such as Mistral Large 4, the models most people can actually run keep getting better in the 7-to-32GB range.

We will add Mellum 2.1 to the model catalogue once we have confirmed the runtime options (particularly official GGUF availability) and can verify its hardware requirements, in line with cataloguing what you can actually run.