One-line: The lightweight variant of the LLaDA2.2 agentic diffusion LLM — 16B total / 1.4B active, Levenshtein Editing (DELETE/INSERT control tokens), 128K context — with real agentic numbers (τ²-Bench 57.50, BFCL-v4 47.68) and a step-change long-context gain (LongBench v2 34.99). ⚠ custom llada2_moe architecture + diffusion runtime (trust_remote_code, SGLang for serving) — no vLLM/Ollama path; all benchmarks lab-reported. Surfaced via lab-org sweep (created 2026-09-05, ~3h outside the 36h cutoff; 0 downloads as of 09-07).
Coding benchmarks
Benchmark
Score
Source
HumanEval
—
not published
HumanEval+
—
not published
MBPP
—
not published
LiveCodeBench
28.14 (v6)
lab (card)
Aider polyglot
—
not published
SWE-bench Verified
—
not published
SWE-bench Multilingual
—
not published
MultiPL-E
65.26
lab (card)
Agentic / tool-use benchmarks
Benchmark
Score
Source
BFCL
47.68 (v4) / 69.02 (v3)
lab (card)
τ-Bench
57.50 (τ²-Bench)
lab (card)
ToolACE
—
not published
GAIA
—
not published
Claw-Eval
57.16
lab (card)
PinchBench
62.33
lab (card)
Reasoning (context only)
Benchmark
Score
Source
MMLU / MMLU-Pro
—
not published
GPQA Diamond
44.41
lab (card; slight dip vs LLaDA2.1-mini 48.36)
MATH
—
not published
AIME 2026
35.05
lab (card)
OlympiadBench
61.11
lab (card)
LongBench v2
34.99
lab (card; up from 12.13 on 2.1-mini)
Efficiency
Active params: 1.4B active / 16B total (incl. embeddings ~; card: 16B excl. embedding) — 256 experts, 8 active per token, 20 layers
Context length: 131,072 tokens (128K native)
VRAM (fp16 rough): ~32 GB (16B × 2) + overhead — fit on a 40-48GB GPU at fp16; realistic local use needs 8-bit/4-bit or a 24GB card with quant
LLaDA2.2-mini is the local-runnable tier of inclusionAI’s agentic diffusion LLaDA2.2 line (flash sibling is ~100B MoE). The interesting signals: Levenshtein Editing (token DELETE/INSERT during diffusion decode) plus L-EBPO agentic RL lift BFCL-v4 from 25.05 (LLaDA2.0-mini) to 47.68, and LongBench v2 jumps 12.13 → 34.99 with the 128K Block Routing — the strongest evidence so far that the post-training, not the base architecture, is what makes the agentic numbers. Caveats: every number is lab-reported with no independent replication, no SWE-bench/Aider published (agent-loop repo-editing suitability unproven), the custom diffusion runtime works only in transformers/SGLang (no llama.cpp — GGUF builds are the fallback, none exist yet), and 0 downloads so far means zero community validation. For a local agent runner it’s a watch-listed experiment, not a daily driver — but it is the only genuinely new lab release with agentic benchmarks in the last ~40h.