LLaDA 2.2 aims at real agents with 1.64× throughput and stronger SWE-bench scores
omarsar0 · x · 2026-07-26
A diffusion LLM for agentic work, LLaDA 2.2 pairs block-parallel decoding with long-horizon planning and tool use.
- The model is presented as the first large-scale diffusion LLM built to act as a real agent, with self-correction over long multi-turn trajectories.
- Benchmark charts show strong results versus Ling-2.6-flash, including SWE-bench Verified 519.0 vs 303.2, SWE-bench Pro 485.3 vs 283.4, SWE-bench Multilingual 459.5 vs 200.6, and τ²-Bench 592.8 vs 334.9.
- The paper also claims 1.64× BF16 throughput over Ling-2.6-flash, with FP8 quantization adding another 18.6%.
- Technical details include native 128K context, MoE routing, Block Routing for predictable long-context cost, Levenshtein editing with KEEP/SUBSTITUTE/DELETE/INSERT, and an RL stage called L-EBPO to reduce trajectory-level error propagation.
More from Models
- Sol61 flunks instruction-following: skips mandated research even when explicitly told — lucasmeijer · 2026-10-26
- Kimi K3.1 Pro tops MathArena open models, closing frontier gap to about 6 months — teortaxesTex · 2026-10-26
- Ruff creator Charlie Marsh: frontier models find infinite bugs in any software — charliermarsh · 2026-10-26
- Dev testing one skill across Opus 5.5 and Sol 6.1 finds instruction following 'all over the map' — lucasmeijer · 2026-10-26
- supermemory unveils learner-1, claiming continual learning for any model — julianweisser · 2026-10-26
- Anthropic says Opus 5.5 cuts cost per task by ~40% vs Opus 5 — taufiqintech · 2026-10-26