NVIDIA releases Deciffusion: 8B diffusion model for 24ms agent decisions
PavloMolchanov · x · 2026-10-17
Pavlo Molchanov (NVIDIA) announced Deciffusion, a decision system finetuned from NVIDIA's Nemotron-Labs-Diffusion-8B diffusion LLM. It returns calibrated probabilities for every option of up to 512 typed questions in a single forward pass, with no generated text.
Key specs:
- 24 ms per decision on one NVIDIA GB200 (NVFP4)
- JevBench public: 210/231; Decision Index 0.3: 53.9 (Tools & Automation 80.2)
- BABILong 32k: 99.6
Trained only on commercially licensed data, quantized to NVFP4 via Model Optimizer's Aumann–Shapley AutoQuantize, and shipped as an open-weights container for agent tool-routing and safety decisions.
More from Models
- Visual reasoning model Elorian Glimpse lands on OpenRouter at $0.60/$3 per 1M tokens — deedydas · 2026-10-17
- NVIDIA releases multimodal reranker on Hugging Face for multilingual vision RAG — solyarisoftware · 2026-10-17
- Lost a bedwars game at 2am, logged into work laptop: users rave 'OS3 will win the agent wars' — SimonBalmain · 2026-10-17
- scaling01: DeepSeek is the Google of China — always late, never frontier — scaling01 · 2026-10-17
- Opus crushes the nacre fan task and extracts signal from DeepSeek's attempts, comparison shows — teortaxesTex · 2026-10-17
- Polymarket odds for 2026's #1 AI model: OpenAI 55%, Meta 16%, Google marked Yes — nian2326076 · 2026-10-17