0xBakeer/TandemLLM-Flint-EXL3-4.0
Text Generation • 10B • Updated • 114
TandemLLM Flint
Recursive's Flint for TandemLLM on one DGX Spark: our EXL3 packs at 4.0 (the default) and 6.0 bits a weight, and Recursive's BF16 checkpoint.
Note The default: 134 tok/s plain on a DGX Spark, KL 0.194 against BF16.
0xBakeer/TandemLLM-Flint-EXL3-6.0
Text Generation • 15B • Updated • 105 • 1Note The quality size: 98 tok/s plain, KL 0.070 against BF16.
RecursiveSI/flint
Text Generation • 40B • Updated • 29.6k • 1.31kNote Recursive's BF16 checkpoint, the reference; TandemLLM serves it at revision 2a247be0.