Mellum2.1 Thinking — GGUF

This repository contains GGUF builds of JetBrains/Mellum2.1-12B-A2.5B-Thinking in every quantization we publish, ready to run with llama.cpp, Ollama, LM Studio, and other GGUF-compatible runtimes. Each variant is one file; download the one you need. The repository also ships the model's multi-token prediction (MTP) head for speculative decoding in llama.cpp; see Speculative decoding with the MTP head.

Mellum2.1 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/Mellum2.1-12B-A2.5B-Thinking.

Available quantizations

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/Mellum2.1-12B-A2.5B-Thinking-GGUF Mellum2.1-12B-A2.5B-Thinking-Q4_K_M.gguf --local-dir .

Run with llama.cpp

# Pull and serve in one step (downloads the GGUF automatically; the suffix picks the quantization)
llama-server -hf JetBrains/Mellum2.1-12B-A2.5B-Thinking-GGUF:Q4_K_M \
  --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 Mellum2.1-12B-A2.5B-Thinking-Q4_K_M.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:

from openai import OpenAI

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

chat_response = client.chat.completions.create(
    model="JetBrains/Mellum2.1-12B-A2.5B-Thinking-GGUF",
    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)

Speculative decoding with the MTP head

Mellum2.1 Thinking was trained with a multi-token prediction head: one extra transformer block (same attention and MoE shape as a model layer) that predicts the token after the one just sampled. llama.cpp uses it as a speculative draft (--spec-type draft-mtp): the full model verifies every drafted token, so the output is unchanged and generation is faster whenever drafts are accepted.

The model weights in the combined files are identical to the quantizations above; in them the head is quantized with the model, and the standalone head file is BF16. These files need llama.cpp build b11747 (19 October 2026) or newer. Older builds, and Ollama, LM Studio, and other runtimes until they ship such a build, cannot load them; they keep working with the quantizations in the first table, which is also what -hf …:Q4_K_M and ollama run hf.co/… resolve to. Without --spec-type draft-mtp nothing changes: the head is simply not used.

# Serve with MTP speculative decoding; llama.cpp also fetches the mtp- draft file from this repository
llama-server -hf JetBrains/Mellum2.1-12B-A2.5B-Thinking-GGUF:Q4_K_M \
  --spec-type draft-mtp --spec-draft-n-max 1 \
  --ctx-size 131072 \
  --temp 0.6 --top-p 0.95 --top-k 20

# Local files: any quantization plus the draft file
llama-server -m Mellum2.1-12B-A2.5B-Thinking-Q4_K_M.gguf \
  -md mtp-Mellum2.1-12B-A2.5B-Thinking-BF16.gguf \
  --spec-type draft-mtp --spec-draft-n-max 1 \
  --ctx-size 131072 \
  --temp 0.6 --top-p 0.95 --top-k 20

# One combined file; the embedded head is the draft
llama-server -m mtp/Mellum2.1-12B-A2.5B-Thinking-MTP-Q4_K_M.gguf \
  --spec-type draft-mtp --spec-draft-n-max 1 \
  --ctx-size 131072 \
  --temp 0.6 --top-p 0.95 --top-k 20

--spec-draft-n-max sets how many tokens the head drafts per step; the head predicts one token ahead, so start with 1. In the llama.cpp benchmarks of this head (Apple M2 Max, Metal, 512 greedy tokens, mean over ten prompts) one drafted token gave 1.05x the BF16 decoding speed, 1.04–1.10x on code prompts, and 1.00x for a Q4_K_M model with a Q8_0 head; two or more drafted tokens were slower than plain decoding in every configuration. The timings object of each server response reports draft_n and draft_n_accepted, the number of drafted and accepted tokens.

Run with Ollama

ollama run hf.co/JetBrains/Mellum2.1-12B-A2.5B-Thinking-GGUF:Q4_K_M

License

Released under the Apache 2.0 license.

---

For the full model card, evaluation results, and architecture details, refer to the original model: JetBrains/Mellum2.1-12B-A2.5B-Thinking.