TensorSharp adds MiniMax-M3.1-Flash support and reports 8× A40 results
A TensorSharp developer reports adding MiniMax-M3.1-Flash support on the first day of its open-weight release, including GGUF inference, dedicated CUDA kernels for MiniMax Sparse Attention, text and image inputs through an OpenAI-compatible API, tool calling, multi-GPU execution, and DSpark speculative decoding. In early tests on eight NVIDIA A40 GPUs, the developer reports 64K executor-prefill throughput of 2,567 tok/s versus 262 tok/s on a tested llama.cpp branch, but says the results came from one host, one quantization, and an uncommitted build without public raw records.
The report provides an early local-inference implementation and a detailed multi-GPU comparison for MiniMax-M3.1-Flash, while its single-host setup and incomplete vision validation limit how broadly the benchmark can be applied.
Some source content is omitted. Continue reading on the original site.
#TensorSharp#MiniMax-M3.1-Flash#NVIDIA A40#llama.cpp#Unsloth#CUDA#DSpark