LLaDA2.2-mini

ADDED 2026-10-17. Open-weight MoE diffusion LM in the sub-8B-active class: card says 16B total excluding embeddings and 1.4B activated. HF API createdAt 2026-09-05T05:16:19Z; the card was lastModified 2026-10-14, which is why it surfaced on this tick. Safetensors total is 16255643392 BF16 parameters, so params_b is 16.256 (the card's 16B is the non-embedding nameplate). active_params_b is the vendor's 1.4B figure, not a count derived here; flops_per_token_active = 2 x 1.4e9 = 2.8e9 is the schema one-pass approximation and does not describe diffusion denoising, which re-reads the block many times. GEOMETRY from config.json fetched 2026-10-17: model_type llada2_moe, hidden 2048, 20 layers, heads 16 / kv heads 4 so gqa_ratio 4, head_dim 128, intermediate_size 5120 for the dense path, moe_intermediate_size 512, 256 experts, 8 per token, num_shared_experts 1, first_k_dense_replace 1 (layer 0 is dense, the rest are MoE), vocab 157184, tie_word_embeddings false, silu, max_position_embeddings 131072. RoPE is native: rope_scaling null, rope_theta 3000000, partial_rotary_factor 0.5, rotary_dim 64, no yarn or ntk factor, so context_scaling is null and context_scaling_method is native. KV: use_cache is false and the card describes iterative block diffusion (block_length 32) plus Levenshtein DELETE and INSERT tokens, so there is no growing autoregressive KV cache. kv_cache_bytes_per_token is 0 for the same reason as llada-8b-instruct. A full bidirectional K+V buffer, if materialised, would be 20 x 4 kv heads x 128 x 2 x 2 = 40960 B per cached token and is not what this field records. AFFORDABLE CLASS: 1.4B active is well under the 8B-active bar and the BF16 checkpoint is about 32 GB, so the weights fit a single consumer GPU. The card's own serving example is still SGLang at tensor-parallel size 8 because diffusion block decoding and the 128K context are the constraint, not the parameter count. No absolute tok/s is published for the mini. The sibling LLaDA2.2-flash card quotes 459-704 TPS against Ling-2.6-flash on agent benches; those numbers are not this model and are not copied here. BENCHMARKS ARE VENDOR CLAIMS on the card, versus LLaDA2.0-mini and LLaDA2.1-mini, not independently re-run: BFCL v4 47.68, AIME 2026 35.05, LiveCodeBench v6 28.14, GPQA-Diamond 44.41, LongBench v2 34.99, general average 46.47, agentic average 59.00 (tau2-Bench 57.50, Claw-Eval 57.16, PinchBench 62.33). Several general scores are below the 2.1-mini row; the card's claim is the agentic lift from Levenshtein editing, not a uniform win. DEPLOYMENT (card): transformers 5.2.0 or newer with trust_remote_code, and SGLang from main (dllm-algorithm JointThresholdInDel, moe-runner-backend triton, example tensor-parallel size 8, context-length 131072). No llama.cpp or Ollama recipe on the card. The technical report is a GitHub PDF (inclusionAI/LLaDA2.X, LLaDA2_2_tech_report.pdf), not an arXiv paper, so no paper source was filed.