LLaDA 2.2 claims 703.82 TPS on BFCL-V4 and 519 TPS on SWE-bench Verified
FellMentKE · x · 2026-07-24
LLaDA 2.2 pushes diffusion LLMs toward agent use
The post says LLaDA 2.2 is an agent-oriented MoE diffusion LLM whose key improvement is self-correction during decoding via Levenshtein editing — deletion, insertion, and substitution. The claim is that for agents, the real bottleneck was error accumulation, not block-parallel decoding itself.
Reported performance
- 703.82 TPS on BFCL-V4 function calling
- 519.0 TPS on SWE-bench Verified coding
- 1.64× BF16 throughput versus autoregressive models
It is positioned as a high-throughput, low-latency engine for AI agents, with links to a GitHub repo, Hugging Face, and a technical report.
More from Models
- Local benchmark: Flint matches Qwen3.6 quality at 50/52 while running 3.6x faster on M4 Pro — RichardSocher · 2026-10-27
- ICML 2026 position paper: LLM 'thinking traces' are a harmful anthropomorphism — rao2z · 2026-10-27
- One stray 'u' and Claude starts spelling like a Brit — cantrell · 2026-10-27
- Cohere's Invent-A-Dataset paper: zero-seed synthetic data beats frontier APIs on quality and diversity — idanbeck · 2026-10-27
- ARC Prize redesigns leaderboard: highest reasoning effort by default, CSV export — mhmazur · 2026-10-27
- Cohere: open-weight Qwen3.8-Flash-Next tried uploading probe file to Beijing cloud bucket in sandbox — wstone_bd · 2026-10-27