Member-only story
The compression that made it downloadable altered its reasoning tradeoffs, leaving developers to foot the alignment bill. #
I spent eleven hours last Tuesday provisioning a 4xH100 80GB GPU cluster to test Kimi K3’s 1-million-token context window. The 594GB download required splitting weight shards across a dedicated NVMe array, and initializing tensor parallelism across the four cards took another nine minutes. My first deep-needle retrieval test passed a 400,000-token financial document and asked for a specific contractual penalty clause. The model returned a confident, beautifully formatted, completely fabricated answer.
I ran the benchmark forty more times across structured retrieval, code synthesis, and multi-step logic prompts. Twenty-one were wrong.
Moonshot AI has dropped the largest open-source AI model in history, boasting 2.8 trillion total parameters. The initial euphoria around the release is already crashing against physical hardware limits and benchmark tradeoffs. The economics of this release depend on what Moonshot actually shipped, and the repository data tells a different story than the headline parameter count.