DreamBlooms/LLaDA2.2-mini-GGUF

LLaDA2.2-mini

LLaDA2.2-mini is the lightweight variant of the agentic diffusion language model in the LLaDA2 series. Built upon the LLaDA2.0-mini architecture, it inherits the core innovations of the LLaDA2.2 series — Levenshtein Editing (introducing DELETE and INSERT control tokens) — enabling long-context tool calling, multi-turn interaction, and robust error correction, while maintaining a smaller parameter footprint and lower inference cost. For more details, please refer to our technical report.

---

📊 Benchmarks

The following table compares LLaDA2.0-mini, LLaDA2.1-mini, and LLaDA2.2-mini across General and Agentic capabilities.

Category	Benchmark	LLaDA2.0-mini	LLaDA2.1-mini	LLaDA2.2-mini
General
Function Calling	BFCL v4	25.05	28.44	47.68
	BFCL v3	70.72	72.06	69.02
Math	AIME 2026	37.71	40.37	35.05
	OlympiadBench	67.70	64.30	61.11
Coding	LiveCodeBench v6	27.70	28.80	28.14
	MultiPL-E	67.46	64.16	65.26
Instruction Following	IFBench	32.33	31.60	24.93
	Multi-IF	60.58	58.43	57.03
Reasoning	KOR-Bench	49.92	46.64	43.60
Knowledge	GPQA-Diamond	47.76	48.36	44.41
Long Context	LongBench v2	15.51	12.13	34.99
General Average		45.68	45.03	46.47
Agentic
Agent	τ²-Bench	-	-	57.50
	Claw-Eval	-	-	57.16
	PinchBench	-	-	62.33
Agentic Average		-	-	59.00

---

🚀 Key Features

• Efficient 128K Diffusion Infrastructure: LLaDA2.2-mini extends the context window to 128K and introduces the Block Routing mechanism, which restricts MoE expert activation at the diffusion block level, enabling efficient long-context agentic tasks.

• Levenshtein Editing: Introduces DELETE and INSERT control tokens, enabling diffusion decoding to edit sequence structure, remove redundant content, and create insertion points during parallel generation.

• Agentic Reinforcement Learning: Proposes Levenshtein Editing ELBO-based Block-level Policy Optimization (L-EBPO), leveraging agentic environment rewards to train Levenshtein editing and error correction capabilities in multi-turn tool-use scenarios.

• Lightweight & Efficient: With a total of 16B parameters and only 1.4B activated during inference, it significantly reduces computational cost while maintaining strong capabilities.

---

📦 Model Variants

Model ID	Description	Hugging Face Link
inclusionAI/LLaDA2.2-flash	Agentic MoE Diffusion Language Model (100B) with Levenshtein editing capabilities.	🤗 Model Card
inclusionAI/LLaDA2.2-mini	Lightweight Agentic MoE Diffusion Language Model (16B) with Levenshtein editing capabilities.	🤗 Model Card

---

🔍 Model Overview

Key specifications of LLaDA2.2-mini:

• Type: Mixture-of-Experts (MoE) Diffusion Language Model with Levenshtein Editing
• Context Length: 128K tokens
• Levenshtein Editing Control Tokens: DELETE, INSERT
• Total Parameters (excl. Embedding): 16B
• Layers: 20
• Attention Heads: 16
• KV Heads: 4
• Experts: 256 (8 activated per token)
• Positional Encoding: Rotary Position Embedding (RoPE)
• Vocabulary Size: 157,184

---

🤗 Hugging Face Transformers Usage

Please ensure transformers>=5.2.0 and related dependencies are installed.

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_path = "inclusionAI/LLaDA2.2-mini"

model = AutoModelForCausalLM.from_pretrained(
    model_path,
    trust_remote_code=True,
    device_map="auto",
)
model = model.to(torch.bfloat16)
model.eval()

tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)

prompt = "Calculate 1+5-28*0.5-200=?"
input_ids = tokenizer.apply_chat_template(
    [{"role": "user", "content": prompt}],
    add_generation_prompt=True,
    tokenize=True,
    return_tensors="pt",
).input_ids

generated_tokens = model.generate(
    inputs=input_ids,
    eos_early_stop=True,
    gen_length=512,
    block_length=32,
    threshold=0.5,
    editing_threshold=0.0,
    temperature=0.0,
)

generated_answer = tokenizer.decode(
    generated_tokens[0],
    skip_special_tokens=True,
)
print(generated_answer)

Best Practices

For optimal performance, we recommend the following configurations:

1. Sampling Parameters: Use block_length=32, temperature=0.0, top_p=None, top_k=None as stable defaults.

2. Denoising Threshold: Adjust threshold, editing_threshold, and max_post_steps based on the speed-quality trade-off for your use case. Lower thresholds can improve inference speed but may lead to repetitive or unstable outputs.

3. Output Length: For most queries, an output length of 32768 tokens is recommended.

4. Long-Context Agentic Tasks: For long-context tool calling and multi-turn agentic applications, we recommend using SGLang as the serving backend. Ensure the server configuration supports a 128K context window and the model's MoE diffusion inference requirements.

---

🖥️ llama.cpp (GGUF)

Pre-converted GGUF quantizations of LLaDA2.2-mini are provided in this repo:

file	quant	size
LLaDA2.2-mini-Q8_0.gguf	Q8_0	17.3 GB
LLaDA2.2-mini-Q4_K_M.gguf	Q4_K_M	9.9 GB

LLaDA2.x is not supported by the stock llama-cli / llama-server yet, and it builds on the DiffusionGemma block-diffusion support (PR #24423), which is not in a llama.cpp release either. Apply both patches to a clean llama.cpp v0.5.0 checkout - first the rebased DiffusionGemma patch, then the LLaDA2 patch:

git clone https://github.com/ggml-org/llama.cpp
cd llama.cpp
git checkout v0.5.0
git apply /path/to/0001-diffusion-gemma-pr24423.patch  # upstream PR #24423, rebased
git apply /path/to/0002-llada2.patch                   # LLaDA2.x support
cmake -B build -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=ON
cmake --build build -j --target llama-diffusion-cli

0001-diffusion-gemma-pr24423.patch is the upstream DiffusionGemma PR #24423 rebased onto the v0.5.0 tag (credit to the DiffusionGemma contributors); 0002-llada2.patch is our LLaDA2.x support on top and does not apply without it.

The patch registers the llada2 architecture, the LLaDA2MoeModelLM converter, and the block-diffusion sampler controls (confidence threshold, EOS early stop, edit-token suppression, active-block prefix). It is adapted from the upstream LLaDA 2.0 diffusion support, PR #17454.

Run a prompt with the recommended decoding (block length 32, threshold 0.5, EOS early stop, greedy):

build/bin/llama-diffusion-cli \
  -m LLaDA2.2-mini-Q8_0.gguf \
  -p "Calculate 1+5-28*0.5-200=?" \
  -ngl 0 -ub 256 -b 256 -c 512 \
  --diffusion-block-length 32 --diffusion-steps 64 \
  --diffusion-threshold 0.5 --diffusion-eos-early-stop \
  -n 256 --temp 0

The chat template is embedded in the GGUF, so -p is wrapped for you. The Levenshtein edit operations (DELETE/INSERT) are not applied by this build; the edit tokens are suppressed so they never leak into the text. Anything that runs these models must go through the llama.cpp diffusion example: the diffusion architectures decode a whole canvas per forward pass and are not driven by the autoregressive loop.

---

🤖 ModelScope

If you are in mainland China, we strongly recommend accessing our models via 🤖 ModelScope.

---

🌐 License

This project is licensed under the Apache License 2.0.

---

🤝 Contact & Collaboration

For any questions, collaboration opportunities, or feedback, please reach out to us via Hugging Face or submit an issue on our GitHub repository.

Join us in advancing open, efficient, and intelligent diffusion language models for agentic applications!

---