Article body wasn't reachable. The HN discussion summary is below.
What commenters are saying
Commenters debate the practicality of the 2.4T parameter model, noting the 1-bit quantized version is 397GB with 95B active parameters per MoE and claims Opus 4.5-level performance. Several users argue this is misleading, as most home labs cannot handle such hardware requirements. Others point out that DeepSeek V4 Flash 0731 offers similar performance at a fraction of the size (284B total, 160GB at FP4). The vision capability is removed and context capped at 250K in the open-source release. A 27B local model is announced for Friday.
What commenters are saying
Commenters debate the practicality of the 2.4T parameter model, noting the 1-bit quantized version is 397GB with 95B active parameters per MoE and claims Opus 4.5-level performance. Several users argue this is misleading, as most home labs cannot handle such hardware requirements. Others point out that DeepSeek V4 Flash 0731 offers similar performance at a fraction of the size (284B total, 160GB at FP4). The vision capability is removed and context capped at 250K in the open-source release. A 27B local model is announced for Friday.