A Dream of Spring for Open-Weight LLMs: 10 Architectures from Jan-Feb 2026

Ahead of AI · 4 months ago

Ten open-weight large language models were released between January and February 2026, including Arcee's Trinity Large (400B parameters), Moonshot's Kimi K2.5 (1 trillion parameters), and StepFun's Step 3.5 Flash (196B parameters), featuring architectural innovations such as sliding window attention, mixture-of-experts designs, and multi-token prediction. Key performance metrics include Step 3.5 Flash achieving 100 tokens/second throughput compared to DeepSeek V3.2's 33 tokens/second, while Qwen3-Coder-Next outperformed larger models on coding benchmarks despite having only 3 billion active parameters. These releases demonstrate ongoing architectural experimentation in open-weight model development, with techniques borrowed from proprietary models being adapted for efficiency and performance across different scales and specializations.