Tag: Mixture of Experts
Inkling 2 Small jumps to 31 on intelligence index for Thinking Machines
The lab's second generation ships as a 594B mixture of experts on Tinker, with open weights waiting on…
Baidu hands out a 280B model with no instructions attached
Baidu published the raw pre-training weights of ERNIE-5.5-Flash-Base on Hugging Face, a 280B mixture-of-experts checkpoint that has never…
Google gives language models a sleep cycle to remember new facts
A consolidation phase moves fragile context into spare expert capacity, holding retrieval up to 10M tokens.
DeepSeek ships a Flash checkpoint and opens its own search API
The 1017 build posts 77.0 on DeepSWE, edging Claude Opus 5 in DeepSeek's own table, while search and…
MiniMax confirms its stealth model was a Flash release all along
Space Bunny spent ten days at the top of OpenRouter's leaderboard before MiniMax revealed it as the new…
Microsoft builds a 2T-parameter base model on a cleaner diet
Microsoft's research group says its new MAI-Base-2 base model beats the best open rivals while using far less…
Meta drops an 816B open model built for coding agents
Muse Spark Open becomes the largest US open-weight release to date, with vision and audio baked in.
Ai2 opens a training stack built for trillion-parameter MoEs
The Allen Institute for AI released Olmo-core 3, an open mixture-of-experts training framework it says scales past a…
Jina AI fits a document parser into a 3.4B open model
The new jina-ocr-v1 turns PDFs, scans and tables into clean Markdown on a single low-cost GPU.
Cohere bets on translation with a 218B open-weight release
North Small Translate arrives with 218B parameters, aggressive per-task pricing and a licence that stops short of commercial…
DeepSeek graduates V4 Pro to production after a four-month preview
DeepSeek's flagship V4 Pro reached general availability as build 0813, ending a four-month preview with unchanged pricing.
AMD’s open 16B MoE model tops fully open rivals on benchmarks
AMD's fully open mixture-of-experts model leads open rivals while activating just 2.8B parameters.