#long-context LLM

'oLLM offloads model weights and KV cache to SSD so you can run very long-context LLMs on a single 8GB GPU, trading storage bandwidth for massive context windows.'

'Alibaba open-sourced Tongyi DeepResearch-30B-A3B, a 30B-parameter MoE LLM tailored for long-horizon, tool-augmented web research with 128K context and strong benchmark scores.'

'Qwen3-Max-Preview is Alibaba's first trillion-parameter LLM, offering a 262K-token context window and competitive benchmark performance while remaining API-only with tiered pricing.'

'NVIDIA's Nemotron Nano 2 delivers hybrid Mamba-Transformer LLMs that run up to 6× faster and support 128K-token context on a single A10G GPU, with most training data and recipes open-sourced.'