DeepSeek V4 Flash drops with 284B parameters

DeepSeek is redefining efficiency by proving that massive parameter counts don't require massive hardware if you're smart about compression and attention.

• –Native FP4/FP8 storage means the "unquantized" weights are already as small as typical 4-bit quants of other models
• –Hybrid Attention architecture reduces the KV cache memory footprint by 90% compared to previous generations
• –With only 13B active parameters, it offers GPT-4 class reasoning speeds on consumer-grade (multi-GPU) or Mac Studio hardware
• –The release of weights under MIT license continues DeepSeek's trend of disrupting the proprietary model landscape
• –"Think" modes allow developers to trade off latency for deeper reasoning depth depending on the task