Xiaomi2026-07-23 23:05:15Xiaomi Launches MiMo V2.5 Multimodal AI, Rivals GPT-5.4 with 42% Lower Token CostXiaomi unveiled MiMo V2.5 and V2.5 Pro, integrating text, image, audio, and video. Pro resolves 57.2% on SWE-bench Pro, matching Claude Opus 4.6 and GPT-5.4, while using 42% fewer tokens than Kimi K2.6. Base model offers lower cost at higher speed, both support 1M context.560
NVIDIA2026-07-23 14:05:15NVIDIA Open-Sources Nemotron 3 Nano Omni, a Multimodal Model for Agentic AINVIDIA released Nemotron 3 Nano Omni, an open-source multimodal model unifying video, audio, image, and text in a single architecture. With 30B-parameter hybrid MoE, it boosts video inference throughput by 9.2x and fully discloses weights, training data, and fine-tuning recipes.560
DeepSeek2026-07-22 15:05:13DeepSeek V4 Specs Leak Claims 1.6T Parameters, 1M Context and Text-Only DesignA post by Princeton AI researcher Yifan Zhang claims DeepSeek V4 will feature 1.6 trillion parameters, a 1 million-token context window and a 285B Lite version, while reportedly remaining text-only.420
MiniMax2026-07-22 08:39:14MiniMax Open-Sources M3: 428B MoE Model with 1M Context and 15x Decoding SpeedupMiniMax open-sourced its flagship M3 model today on Hugging Face. It features a 428B MoE architecture with only 23B active parameters per token, supports 1M token context, and achieves up to 15x decoding speedup via its proprietary MSA sparse attention mechanism.510
Meta2026-07-11 01:00:13Meta Multimodal Model Sweeps Seven Benchmarks, GPQA Diamond Hits 89.5%Meta's latest multimodal model achieved top rankings across seven major AI benchmarks, including 89.5% on GPQA Diamond and 42.5% on ARC-AGI-2, marking a significant comeback in the AI race.380
