MON, AUGUST 03, 2026
Independent · In‑Depth · Practitioner‑Tested
Large Language Models

Qwen3.8-Max-Preview vs Kimi K3 (2026): 2.4T Preview vs 2.8T Open Weights

Two Chinese 2T+ Frontier Models — One Has Open Weights and Published Benchmarks, One Does Not

🕐 5 min read 👁 11 views 📅 Aug 3, 2026

QUICK VERDICT — AUGUST 2026

Kimi K3 now: Use Kimi K3 for production 2T+ frontier work. Self-hosted via Modified MIT weights on Western cloud reduces the Treasury sanctions risk. Freeze new Moonshot API commitments until the sanctions situation resolves.

Test Qwen3.8-Max-Preview: Run via Token Plan at ~10% of standard pricing. When Alibaba publishes a benchmark table, active parameter count, and open-weight checkpoint — reassess. If the claims hold, it could match or beat K3 without the regulatory risk.

Last updated August 3, 2026. Related: Qwen3.8-Max full fact-check → · Kimi K3 download guide →

⚖ Our Verdict

Kimi K3 wins on every verifiable dimension today — published benchmarks, live open weights, confirmed API price. Qwen3.8-Max-Preview has none of these. Test Qwen3.8 via Token Plan preview; use Kimi K3 (self-hosted via Modified MIT on Western cloud to reduce sanctions risk) for production work needing a 2T+ frontier model now.