103B total (100B non-embedding, MoE)
128K tokens
Apache 2.0
Agent-oriented Mixture-of-Experts diffusion language model designed for long-context tool use, multi-turn interaction, and error correction. LLaDA2.2 extends the series from fixed-length token substitution to Levenshtein Editing with KEEP, SUBSTITUTE, DELETE, and INSERT operations, allowing parallel diffusion decoding to remove redundant spans, insert missing positions, and change sequence length. Its L-EBPO reinforcement-learning method combines longest-common-subsequence-derived editing labels with environmental rewards to train editing and recovery behavior during long-horizon agent trajectories. The 32-layer architecture uses 256 routed experts with eight selected per token plus one shared expert, 32 attention heads, a 4,096-dimensional hidden state, RoPE, a 157,184-token vocabulary, and 32-token diffusion blocks. Progressive long-context continual pretraining extends the model to 128K, while Block Routing regulates expert activation at the diffusion-block level to control MoE inference costs. Official evaluations report 80.33 on tau2-Bench, 81.66 on PinchBench, 64.22 on Claw-Eval, 46.21 on MCP-Atlas, 49.28 on SWE-bench Verified, and 60.78 on BFCL-V4. The technical report reports an average 1.64x BF16 throughput advantage over Ling-2.6-flash across 11 tested workloads, including 703.82 tokens per second on BFCL-V4; FP8 quantization provided a further 18.6% average throughput increase. The release supplies approximately 206 GB of BF16 safetensors and custom Transformers implementation code; the model card notes that dedicated SGLang deployment support is still forthcoming.