LLaDA2.0-flash (100B-A6B)
LLaDA2.0-flash — a 100B-A6B MoE diffusion language model, the instruction-tuned flagship of the LLaDA2.0 series (alongside LLaDA2.0-Uni and LLaDA2.0-mini), averaging 79.32 across knowledge, reasoning, coding, math, agent and alignment benchmarks.
Overview
LLaDA2.0 is a family of discrete diffusion large language models (dLLMs) scaling up to 100B total parameters, established through systematic conversion from auto-regressive (AR) models rather than costly training from scratch. It follows three design principles: knowledge inheritance, progressive adaptation, and efficiency-aware design.
The flagship LLaDA2.0-flash is a 100B-A6B MoE diffusion model that reaches an average score of 79.32 across knowledge, reasoning, coding, math, agent, and alignment benchmarks — competitive with contemporary dense and MoE AR models at a fraction of the inference cost. The series also includes LLaDA2.0-Uni, a unified dLLM supporting multimodal understanding and generation, and LLaDA2.0-mini for lightweight deployments.
LLaDA2.0-flash reaches ~79.3 average across benchmarks, on par with Qwen3-30B-A3B-Instruct-2507 and Ling-flash-2.0 while remaining a diffusion model.
Model Series
| Model | Params | Notes |
|---|---|---|
LLaDA2.0-flash | 100B-A6B | Flagship instruction-tuned MoE dLLM |
LLaDA2.0-mini | small | Lightweight variant |
LLaDA2.0-Uni | MoE | Unified multimodal understanding & generation dLLM |
LLaDA2.0-flash-preview / mini-preview | — | Preview releases |
LLaDA2.0-flash-CAP / mini-CAP | — | Specialized variants |
Highlights
- 100B-scale diffusion: first open dLLM family at 100B scale with MoE sparsity (flash: 100B total / 6B activated).
- Knowledge inheritance from AR models: converted from auto-regressive checkpoints, preserving prior knowledge while adding diffusion capabilities — no from-scratch training at frontier scale.
- Strong general abilities: MMLU 87.69, HumanEval 94.51, GSM8K 96.06 on flash; comparable to leading MoE models of similar activated size.
- Multimodal unification: LLaDA2.0-Uni natively supports multimodal understanding and generation within one diffusion LLM.
- Ecosystem: released together with the dFactory training framework and inference tooling from inclusionAI.