UPDATED JULY 27 — BOTH NOW SELF-HOSTABLE
● K3 weights: LIVE — huggingface.co/moonshotai/Kimi-K3, Modified MIT
● Better intelligence: Kimi K3 — AA Index #4 (score 57) vs V4 Pro not ranked at frontier
● Better agentic coding: Kimi K3 — SWE Marathon 42.0% #1
● Lower VRAM (easier deploy): DeepSeek V4 Pro — significantly smaller than 2.8T K3
● More production-mature tooling: DeepSeek V4 Pro — months of community vLLM/quant work
● Hallucination warning: K3 only — 51% rate from independent testing not in Moonshot benchmarks
● Data residency (self-hosted): Both resolved on Western cloud
Self-hosted K3 for: Teams that need frontier-tier intelligence and the strongest agentic coding from an open model. Worth the VRAM overhead (18+ H100 80GB at Q4) for workloads where K3's benchmarks matter. Test hallucination on your use case first.
Self-hosted V4 Pro for: Lower VRAM, deploy today (months of community tooling), proven production stability, and workloads where K3's hallucination risk is unacceptable. 7x cheaper than K3 on API if staying managed.
Last updated July 27, 2026. Related: K3 download guide → · DeepSeek V4 Flash vs V4 Pro →
⚖ Our Verdict
Kimi K3 wins on intelligence (AA Index #4), agentic coding (SWE Marathon #1), and context (1M). DeepSeek V4 Pro wins on VRAM requirements (far smaller, easier deploy), production maturity (months of tooling), and hallucination safety (K3 has 51% rate from independent testing). Both now self-hostable on Western cloud — China NI Law resolved for both.