The Number
[MODEL] Moonshot's kimi-k3-max ranks #2 on LMArena's coding preference board at Elo 1682 (2026-07-29), behind only claude-opus-5-max. Preference ranking, not capability.
Untranslated
Read the coding leaderboard, not the overall one, if you want to see where China's labs actually sit. On LMArena's general text preference board (as of 2026-07-29), one Chinese model cracks the top twelve: kimi-k3-max at #11, Elo 1486, tied with Google's gemini-3-pro. The vision board returns no China entry in its top slice at all. But the coding board tells a different story: kimi-k3-max is #2 at Elo 1682, Z.ai's glm-5.2-max is #6 at 1588, and below them a long dense tail no Western lab matches for depth. Count the China entries on the code board: Moonshot, Z.ai, ByteDance's seed-2.1-pro-preview (#17), Tencent's hy3 (#20), Alibaba's qwen3.7-max (#21), MiniMax m3, Xiaomi's mimo, DeepSeek's v4-pro-thinking (#33). Nineteen-plus China-lab models in the returned code slice against a handful on the text board. All of this is human-preference Elo, gameable by answer style, and new models take days to stabilize, so treat the exact ranks as soft. The hard signal underneath them is price. On OpenRouter (developer-heavy user base, not global usage) the China-lab column reads like a floor collapse: qwen3.7-flash at $0.03 per million prompt tokens, glm-4.7-flash at $0.06, Tencent hy3-preview at $0.063, ByteDance seed-1.6-flash at $0.075. The vendor pages confirm it is not a routing artifact. DeepSeek's own docs (as of 2026-07-30) list deepseek-v4-flash at $0.14 per 1M input on cache miss, $0.28 output, and a cache-hit input price of $0.0028. Qwen's DashScope page (2026-07-30) shows qwen3-max at 2.5 yuan per 1M input in the smallest tier. Moonshot's platform now leads with Kimi K3 as flagship and marks the old Moonshot V1 series for full retirement around 2026-08-31. The pattern to carry away: China's labs are not fighting for the single top Elo seat that Anthropic owns on every board here. They are stacking a wide, coding-competent, radically cheap tier and moving customers onto it. The coding board is where that strategy shows up as a rank; the price columns are where it shows up as a business.
The Kicker
[MONEY] Buried in a 36氪 morning digest (2026-07-29): Moonshot's Kimi has closed an over-$3.5B Series F. The company that just retired its first-ever model series is raising like the tier war is only starting.