Minima Research
Minima research focuses on reducing LLM inference memory and serving cost while preserving practical deployment constraints.
Paper
Minima shows that the recurrent half of a hybrid 27B LLM can be quantized to 4-bit NVFP4 without meaningful quality loss. On Qwen3.8-27B, the resulting 17.5 GiB model delivers 14–19% higher token throughput than comparable quantized baselines.
Read paper (opens in a new tab)Paper
Minima-KV introduces mixed-format paged attention for compressing long-context KV cache without evicting live-request pages. On Qwen3.6-27B, it achieves 3.50× compression versus BF16 while preserving near-baseline retrieval quality.
Read paper (opens in a new tab)Paper
Minima combines tensor-network compression, healing and custom GPU kernels to reduce the memory footprint of production-scale LLM inference. On Qwen3-32B, it reduces peak VRAM from 64 GiB to 40 GiB and increases single-request throughput from 40 to 50 tokens per second.
Read paper (opens in a new tab)