Released my custom SGLang V14 runtime for persistent cross-request LLM memory. Tested with Qwen3.8-27B and Nemotron 120B (text requests). Qwen4-Exp is a future research target.
Docker: https://t.co/4UWTV3NStO
More models: https://t.co/hkDrYUTHAS
#AIResearch
V14 Episodic Memory is out for Qwen3.8-27B and Nemotron 3.5 Super VL 120B (NVFP4). Both have memory-enabled tests; Nemotron validation covers text requests.
Qwen: https://t.co/mj3DoOy41U
Nemotron: https://t.co/VHl5siuL50
V14 Episodic Memory is out for Qwen3.8-27B and Nemotron 3.5 Super VL 120B (NVFP4). Both have memory-enabled tests; Nemotron validation covers text requests.
Qwen: https://t.co/mj3DoOy41U
Nemotron: https://t.co/VHl5siuL50
#OpenSourceAI#LLM
Independent AI researcher seeking a donated DGX Station GB300 (or equivalent high-memory system) for persistent-memory LLM research. V14 tested on 27B and 120B models. Preparing for Qwen4-Exp. My-Models: https://t.co/hkDrYUTHAS
Independent AI researcher seeking a donated DGX Station GB300 (or equivalent high-memory system) for persistent-memory LLM research. V14 tested on 27B and 120B models. Preparing for Qwen4-Exp.
My-Models: https://t.co/hkDrYUTHAS
My-Code: https://t.co/T8Xev8R6z0
#AIResearch
V14 Episodic Memory is out for Qwen3.8-27B and Nemotron 3.5 Super VL 120B (NVFP4). Both have memory-enabled tests; Nemotron validation covers text requests.
Qwen: https://t.co/mj3DoOy41U
Nemotron: https://t.co/VHl5siuL50
#OpenSourceAI#LLM
@teortaxesTex it better not be moe istg moe is not useful unless youre in the 100+B range the only reason 3.8-27B is so good is because it uses all the parameters it can get its hands on