LLaDA2.2 targets the real bottleneck in multi-turn agents: decode speed
Direct_Band896 · reddit · 2026-07-25
Decode speed is the hidden tax in agent loops
A Reddit post argues that the real bottleneck in multi-turn agent systems is not tool calls or retries, but raw token-by-token decoding. In agent workflows, the model has to decode repeatedly across planning and tool-use turns, so serial generation compounds both latency and serving cost.
LLaDA2.2 aims at that bottleneck
The post highlights LLaDA2.2, a 100B-class MoE diffusion model built for multi-turn tool use. Instead of locking tokens in place one by one, it decodes blocks and can keep, substitute, delete, or insert tokens mid-decode.
Reported numbers
- About 1.6x average decode throughput vs. its autoregressive sibling
- Up to 2.3x on agent workloads
- Around 703 tokens/sec on the BFCL function-calling set
- Better interactive-agent results than the AR sibling:
- tau2 bench: 80.33 vs 76.36
- MCP Atlas: 46.21 vs 41.12
Caveats
It is not a general reasoning upgrade: the post says it trails the AR sibling on SWE-style benchmarks and general knowledge tasks. It also notes practical deployment limits today: 205.8 GB weights, Apache 2.0, no llama.cpp path, and SGLang support only coming soon.
The core takeaway is that if agent latency is dominated by serial decoding, block-level diffusion decoding is pointed directly at the tax teams are actually paying.
More from coding & agent
- 'Let an agent run wild on vague goals and it makes your code worse 9 out of 10 times' — eptwts · 2026-10-27
- LangChain to demo the in-house design agent its team uses daily, this Wednesday — LangChain · 2026-10-27
- Biotech Agent Series Day 8: A Notebook Review Agent to Auto-Audit Lab Records — nlarusstone · 2026-10-27
- Photon Launches Agent-to-Agent Messaging in iMessage for Real AI Collaboration — alvelda · 2026-10-27
- Fake-paper meme shows an agent workflow for reverse engineering trading systems into a one-person quant fund — PtrPomorski · 2026-10-27
- Codex clones Linear's core from a one-page spec in 3 minutes 28 seconds — aakashgupta · 2026-10-27