tinyanalysis

Leaderboard / JetBrains (Mellum)

Mellum2.1 12B-A2.5B Thinking

JetBrains (Mellum) · Mellum2 · released 2026-09-20 · data checked 2026-11-04

Mellum2.1 12B-A2.5B Thinking is a 12.2B-parameter Mixture-of-Experts model (2.5B active per token) from JetBrains (Mellum) (Europe), released 2026-09-20 under the Apache-2.0 license. It scores 60.8 on the Small LM Index, #5 of 30 small models. It needs about 8.3 GB of memory at 4-bit with an 8K-token context.

JetBrains' own Mixture-of-Experts model for agentic coding (working in a repository, running commands, calling tools), post-trained mainly with reinforcement learning from Mellum2 Base.

Small LM Index
60.8
#5 of 30 scored small models
Likely 58.4–63.1 · from 4 of 6 core benchmarks
Parameters
12.2B
2.5B active · MoE
Price per 1M tokens
—
No hosted API on OpenRouter; self-host
Context window
131K
tokens

Benchmarks

Each bar shows this model's score. The tick marks the best small model on that benchmark. Sources: 8 reported by JetBrains (Mellum). How we label sources.

GPQA Diamond Lab-reported Knowledge · GPQA Diamond64.6
Class best: Qwen3.5 9B 80.6
AIME Lab-reported Math · Mean of AIME 2025 and 202683.3
Class best: Apriel 1.6 15B Thinker 88.0
LiveCodeBench Lab-reported Coding · LiveCodeBench v682.0
SWE-bench Verified Lab-reported Coding47.0
Terminal-Bench Lab-reported Agentic · Terminal-Bench 2.117.4
Class best: Granite 4.2 8B 20.6
MMLU Lab-reported Knowledge · MMLU-Redux87.8
IFEval Lab-reported Instruction following90.6
Class best: LFM2.5 8B-A1B 91.8
BFCL Lab-reported Agentic · BFCL v4 (macro-average)62.3
Class best: Hunyuan 7B Instruct 70.8

Not reported: estimated

What the index model expects from this model's other scores. Not a measurement.

MMLU-Pro≈ 79
Humanity's Last Exam≈ 10

Specifications

Lab
JetBrains (Mellum) (CZ · Europe)
Architecture
Mixture-of-Experts
Total parameters
12.2B
Active parameters
2.5B
Size class
Small
Context window
131K tokens
Input modalities
text
Reasoning
Reasoning (thinking)
License
Apache-2.0
Openness
Open weights
Memory to run
~8.3 GB at 4-bit, 8K context (more sizes)
Hugging Face
10.5K downloads (30 days) · 190 likes
Weights on Hugging Face ↗

Compare with

Ranked in

Sources

Something wrong here? Report it.

Run it locally

Uses JetBrains/Mellum2.1-12B-A2.5B-Thinking-GGUF (Q4_K_M). Swap the tag for another size: BF16, MXFP4, Q4_K_M, Q6_K, Q8_0.

Ollama
ollama run hf.co/JetBrains/Mellum2.1-12B-A2.5B-Thinking-GGUF:Q4_K_M
llama.cpp
llama-server -hf JetBrains/Mellum2.1-12B-A2.5B-Thinking-GGUF:Q4_K_M
Then open http://localhost:8080 for a chat window.
MLX (Mac)
pip install -U mlx-lm
mlx_lm.chat --model mlx-community/Mellum2.1-12B-A2.5B-Thinking-6bit
vLLM (GPU server)
vllm serve JetBrains/Mellum2.1-12B-A2.5B-Thinking
Full-precision weights, for a server GPU.

Open in LM Studio

Memory needed: weights, context cache and runtime buffers.

Quantization4K context32K context131K context
16-bit F1625 GB26 GB27 GB
8-bit Q8_014 GB14 GB16 GB
4-bit Q4_K_M8.2 GB8.6 GB10 GB

Check it against your GPU, Mac or phone →