DeepSeek V4: Open-Weight Frontier at 1/7 the Cost
May 2, 2026
DeepSeek V4 ships 1.6T MoE open weights with 1M-token context: 80.6% on SWE-bench Verified at $1.74/$3.48 per million — roughly 1/7 the output cost of Claude Opus 4.7.
#mixture-of-experts
Kimi K2.6: Open-Weight 300-Agent Swarm Tops GPT-5.4 (2026)
April 27, 2026
Moonshot AI's Kimi K2.6 (April 20, 2026): a 1T-parameter open-weight MoE that orchestrates 300 sub-agents and tops GPT-5.4 on SWE-Bench Pro at 58.6%.
Gemma 4 Review: Apache-2.0 Open AI, 89.2% on AIME 2026
April 4, 2026
Google DeepMind's Gemma 4 family: 31B Dense scores 89.2% on AIME 2026, 26B MoE activates 3.8B/token, E2B fits in 1.5 GB. Apache 2.0, runs on your GPU.
Qwen3.5-Omni: Alibaba Omnimodal AI Model (2026)
April 1, 2026
Alibaba's Qwen3.5-Omni processes text, images, audio, and video natively with real-time speech output, 113-language recognition, and 215 SOTA audio subtasks.
DeepSeek V3 Coding: Power, Pricing, and Practical Integration
March 3, 2026
DeepSeek V3 coding: 671B MoE, 82.6% HumanEval, beats GPT-4o and Claude 3.5 on 5 of 7 coding benchmarks. Pricing, integration patterns, and real caveats.