Kobold
An open-source inference backend tool for running large language models locally. It is commonly used in role-playing and creative writing scenarios and supports multiple quantization formats as well as context length extension.
Timeline (last 90 days)
The hardware baseline for the test was 10–16 GB VRAM combined with 26–32 GB of system memory, which, using a Kobold or llama.cpp backend, can run at up to a 64K context window.
Kobold cannot load larger models; Qwen 35B (approximately 20.6 GB) and GLM 31B (approximately 17.7 GB) require the LLAMA backend.
All Facts (2)
The hardware baseline for the test was 10–16 GB VRAM combined with 26–32 GB of system memory, which, using a Kobold or llama.cpp backend, can run at up to a 64K context window.
50%UnverifiedKobold cannot load larger models; Qwen 35B (approximately 20.6 GB) and GLM 31B (approximately 17.7 GB) require the LLAMA backend.
50%