This repository contains a GGUF MXFP4_MOE quantization of JetBrains/Mellum2-12B-A2.5B-Thinking, ready to run with llama.cpp, Ollama, LM Studio, and other GGUF-compatible runtimes.

This quantization (MXFP4_MOE): MXFP4 microscaling 4-bit applied to the MoE expert tensors. Smallest footprint, with a modest quality cost (KLD ~0.088, 87% top-token agreement).

File	Size
`Mellum2-12B-A2.5B-Thinking-MXFP4_MOE.gguf`	7.0 GB

Mellum 2 Thinking is a Mixture-of-Experts reasoning model (64 experts, 8 activated per token, 131,072-token context) that emits its chain of thought inside <think>...</think> blocks before the final answer. For the full model description, evaluation results, and architecture details, see the original model card: JetBrains/Mellum2-12B-A2.5B-Thinking.

Available quantizations

Quantization	Description	Size	KLD vs BF16 ↓	Top-token match ↑
`BF16`	16-bit, no quantization (reference)	24.3 GB	—	—
`Q8_0`	8-bit, effectively lossless	12.9 GB	0.004	97.4%
`Q6_K`	6-bit k-quant, very high quality	10.9 GB	0.014	95.1%
`Q4_K_M`	4-bit k-quant, balanced (recommended)	8.1 GB	0.052	89.8%
`MXFP4_MOE` (this repo)	MXFP4 4-bit on MoE experts, smallest	7.0 GB	0.088	87.3%

KL divergence and top-token agreement are measured against the BF16 logits on Wikitext-2 (n_ctx=512); lower KLD / higher agreement means closer to the unquantized model.

Download

hf download JetBrains/Mellum2-12B-A2.5B-Thinking-GGUF-MXFP4_MOE Mellum2-12B-A2.5B-Thinking-MXFP4_MOE.gguf --local-dir .

Run with llama.cpp

# Pull and serve in one step (downloads the GGUF automatically)
llama-server -hf JetBrains/Mellum2-12B-A2.5B-Thinking-GGUF-MXFP4_MOE \
  --ctx-size 131072 \
  --temp 0.6 --top-p 0.95 --top-k 20

# Or run a one-off prompt with a local file
llama-cli -m Mellum2-12B-A2.5B-Thinking-MXFP4_MOE.gguf \
  --ctx-size 131072 \
  --temp 0.6 --top-p 0.95 --top-k 20 \
  -p "Is 1024 a power of 2? Explain your reasoning."

The server exposes an OpenAI-compatible API on http://localhost:8080/v1:

python

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8080/v1", api_key="llama.cpp")

chat_response = client.chat.completions.create(
    model="JetBrains/Mellum2-12B-A2.5B-Thinking-GGUF-MXFP4_MOE",
    messages=[
        {"role": "user", "content": "Is 1024 a power of 2? Explain your reasoning."},
    ],
    max_tokens=81920,
    temperature=0.6,
    top_p=0.95,
    extra_body={"top_k": 20},
)
print(chat_response.choices[0].message.content)

Run with Ollama

ollama run hf.co/JetBrains/Mellum2-12B-A2.5B-Thinking-GGUF-MXFP4_MOE

License

Released under the Apache 2.0 license.

For the full model card, evaluation results, and architecture details, refer to the original model: JetBrains/Mellum2-12B-A2.5B-Thinking.