inclusionAI Launches LLaDA2.2-flash: 100B Diffusion Model for Agent Acceleration

inclusionAI released LLaDA2.2-flash, a 100B MoE diffusion language model designed for agentic workloads. Featuring a native 128K context window, it aims to improve multi-turn tool calling and long-context interactions through parallel decoding, offering a novel approach to reducing agent latency.

Confirmed

The model is a 100B MoE with a native 128K context window. Its core mechanism utilizes Levenshtein editing (including KEEP, SUBSTITUTION, DELETE, and INSERT operations) to achieve parallel decoding. In agent evaluations, according to @FellMentKE and @qruiq, the model delivered high throughput on BFCL and SWE-bench, with inference speeds reaching up to 2.3 times that of traditional autoregressive models. @heyshrutimishra also noted a 1.6x overall speedup in agentic workloads.

Unconfirmed

The overall capabilities of diffusion models still face limitations. @Additional-Engine402 pointed out through same-scale comparative evaluations that LLaDA2.2 lags behind autoregressive models in general knowledge and most coding benchmarks, currently holding a speed advantage only in a few agent scenarios.

Why it matters

Diffusion LLMs are expanding into complex agentic applications. @DirectBand896 noted that in multi-turn agent tasks, the true bottleneck slowing down response and driving up inference costs is often the model's token-by-token serial decoding itself, rather than the number of tool calls. By introducing edit-based diffusion and parallel decoding, LLaDA2.2-flash directly addresses this latency pain point, providing a new solution for high-throughput, long-context agent inference.

2026-07-23 ~ 2026-07-25 · 9 related posts

Primary sources