Reka AI's Flash 3.1 release pairs a 21-billion-parameter LLMA large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.Full definition → with a compressed version and the Reka Quant library. Compression stores the model's learned numerical weights using fewer bits, reducing its memory footprint. The release includes a version using approximately 3.5 bits per weight.
The technical account, also covered in the companion Intel entry, explains how it tries to preserve capability: progressively compress weights while training the remaining model to match the original higher-precision model. Reka reports an average 1.6-point decline on its hard-reasoning benchmarks for that compressed version, making the memory reduction an explicitly measured trade-off.
Terms in this piece · Glossary
LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.
distillation — Training a small, cheap model to imitate a big one's outputs, keeping much of the capability at a fraction of the cost.
quantization — Shrinking a model by storing its numbers less precisely — like rounding — so it runs faster and fits on smaller hardware, at a small quality cost.
