Robots Atlas>ROBOTS ATLAS

Tag: Speculative decoding

MiniMax M3: sparse attention architecture and 15.6× faster decoding
Aktualności28 maja 2026

MiniMax M3: sparse attention architecture and 15.6× faster decoding

MiniMax published a technical report on its M2 series and announced M3 — a model with a new sparse attention mechanism (MSA) that decodes 15.6 times faster than M2 at one-million-token context lengths. It is the first sub-quadratic architecture the company says preserves multi-hop reasoning without compromise.

Cerebras Runs Trillion-Parameter AI Model Nearly 7x Faster Than GPU Clouds
Aktualności26 maja 2026

Cerebras Runs Trillion-Parameter AI Model Nearly 7x Faster Than GPU Clouds

Cerebras Systems announced 981 output tokens per second for Kimi K2.6 — 6.7 times faster than the next GPU-based cloud provider. Result independently verified by Artificial Analysis. The announcement came less than a week after the largest tech IPO of 2026.

NVIDIA Nemotron-Labs-Diffusion: AR and diffusion in one model
Aktualności22 maja 2026

NVIDIA Nemotron-Labs-Diffusion: AR and diffusion in one model

NVIDIA released the Nemotron-Labs-Diffusion model family (3B/8B/14B) — the first production-grade implementation combining autoregression with block-wise diffusion decoding and a new Self-Speculation mode that achieves 3.3× higher throughput than classical AR on GB200.

Google boosts Gemma 4 inference up to 3x with speculative decoding
Aktualności6 maja 2026

Google boosts Gemma 4 inference up to 3x with speculative decoding

On May 6, 2026, Google released experimental Multi-Token Prediction (MTP) drafter models for the Gemma 4 family, accelerating local inference up to three times with no loss of output quality. The technique is based on speculative decoding: a lightweight draft model predicts future tokens, which are then verified in parallel by the main model.