Side-by-side eval shows diffusion loses overall, but wins speed in agent loops
Additional-Engine402 · reddit · 2026-07-25
The post breaks down a new side-by-side comparison between a diffusion model, LLaDA2.2, and a same-size autoregressive model from the same lab.
- On general knowledge and most coding benchmarks, the diffusion model underperforms.
- It does beat the autoregressive sibling on a few agent-style, multi-turn tool-use tasks, including tau2 bench (80.33 vs 76.36) and MCP Atlas (46.21 vs 41.12).
- The authors say the SWE-bench gap is hard to read literally because the two runs used different scaffolds.
- The biggest win is speed: about 1.6× average decoding throughput with BF16 + speculative decoding, and up to 2.3× on agent workloads.
- The weights are open under Apache 2.0, but the download is huge (205.8 GB), there is no llama.cpp support, and serving is only “coming soon.”
The post’s conclusion is narrow: diffusion does not “win” overall, but it may be a better backbone for agent loops where decode latency matters every turn.
More from Models
- Kimi K3.1 Pro regresses on Debate Benchmark while DeepSeek 1017 gains at lower cost — teortaxesTex · 2026-10-27
- 4 Hours of LLM Startup Research Produced a Convincing but Nonsense Product Thesis — annetgriffin · 2026-10-27
- Dev argues Claude burns far more tokens than GPT models for the same coding work — haider1 · 2026-10-27
- MiniMax matches Claude Opus 4.5 in non-benchmaxxed evals, users report — teortaxesTex · 2026-10-27
- Real-world math: GPT-6.1 Terra's half price is offset by 2-3x token usage, costing more than Sol — lordpuddingcup · 2026-10-27
- Fable 5.5 rumored to ship this week as partners still test — kimmonismus · 2026-10-27