Kimi K3.1 Pro: Your Quota Is Depleted | Moonshot AI | Kimi Subscription Limits | AI Model Review ...

Original Media · 17m · 中文 · YouTube · Best Partners TV

TLDR

Moonshot AI's Kimi K3.1 Pro launch showcases strong technical gains via post-training RL — hitting near-SOTA coding/agent benchmarks without parameter scaling — but is overshadowed by severe quota restrictions (5-hour windows exhausted in minutes), high subscription prices (199/699 RMB), a covert marketing scandal overseas, and fundamental compute scarcity from US chip sanctions. Despite $1B+ ARR, the model sits awkwardly between cheaper domestic rivals (DeepSeek, MiniMax) and more capable foreign flagships (Claude, GPT), with token price cuts offset by longer agentic reasoning.

For you

PRO members can sign in and request more information — or their own version of this digest, tuned to what they care about.

Sign in to unlock

Surprising Insights

•

Despite $1B ARR and 20-day revenue exceeding all of 2025, Moonshot still cannot serve enterprise demand due to GPU shortage — same 'compute not enough' excuse from 2024.

•

K3.1 Pro's average task cost equals K3 despite 33% lower token prices, because the model 'thinks longer, uses more tools, retries more' — efficiency gains eaten by agentic behavior.

•

The 65.7% Terminal-Bench score matches Claude Opus 5.5 (66.4%) but is self-reported; third-party Artificial Analysis shows Kimi at 67% vs Sonnet 5.5's 64%, but test setups differ.

•

Chinese model labs' 'native family tragedy': US sanctions forced efficiency culture (DeepSeek's MoE/MLA), but open-sourcing weights lets cloud providers capture inference revenue, leaving originators with R&D costs but no runtime margin.

•

Huawei Ascend 950DT (144GB, 4TB/s) nears H200 specs; DeepSeek reportedly planning 160k Ascend 950DT deployment — domestic compute scaling fast but orders allegedly booked to 2030.

Key takeaways

•

K3.1 Pro achieves 65.7% Terminal-Bench and 61.2% Kimi Code Bench via massive post-training RL on 2.8T MoE base (104B active params), no parameter increase.

•

User backlash over quotas: 5-hour windows depleted in minutes for complex tasks; 699 RMB plan delivers ~1.94x less API value than predecessor.

•

Marketing scandal: Overseas creators asked to hide paid promotion tags, violating FTC/platform rules, damaging credibility despite model quality.

•

Pricing paradox: Token prices down 33% vs K3, but average task cost unchanged (~$1.98) due to 14% more output tokens from extended reasoning/tool use.

•

Compute scarcity is root cause: US chip sanctions limit inference capacity, forcing quota rationing despite revenue growth.

•

Domestic rivals (DeepSeek Flash, MiniMax M3.1 Flash) offer 8-16x cheaper inference via architectural efficiency (MoE, MLA, Flash models).

•

K3.1 Pro only supports mandatory thinking mode (low/high/max), cannot disable reasoning, inflating token usage for simple tasks.

•

Open weights promised by Nov 4; safety measure: removed cyberattack data from open weights, reducing exploit success from 90.5% to 81%.

Notable quotes

“思考中,思考中,思考中,您的额度已用完”

▸ Notable quotes · 0:21

“说发现「可能」之极限的唯一方法 是向「不可能」的方向多走一小步”

▸ Notable quotes · 1:25

“这就像尽全力培养了一个高材生 智商很高,但是一考试就晕倒了”

▸ Notable quotes · 2:26

“单价降了账单未必降”

▸ Notable quotes · 8:33

“就是算力不够”

▸ Notable quotes · 11:18

“过去关心一个Token多少钱 如今更该关心完成一项工作要消耗多少计算资源”

▸ Notable quotes · 12:21

Chapters

•

0:00 — Launch Backlash: Quota Crisis & User Anger: K3.1 Pro launch dominated by complaints: 199/699 RMB plans offer 5-hour windows exhausted in minutes; users report 699 plan gives ~1.94x less API value than old plan; Reddit users hit limits on basic 3D coding tasks.

•

2:47 — Marketing Scandal: Covert Paid Promotions: Overseas creator (6k followers) exposes Moonshot's requirement to hide paid promotion tags, skip official links/mentions — violating FTC/platform rules; community criticizes unnecessary deception for a strong model.

•

3:29 — Model Deep Dive: Architecture, Benchmarks & Capabilities: K3.1 Pro: 2.8T MoE (104B active), 1M context, post-training RL across hundreds of domains. Scores: Terminal-Bench 65.7% (near Opus 5.5), Kimi Code Bench 61.2% (+5.3pp), hallucination rate 8.3% (down from 23.4%). Strong in 3D parametric modeling (Blender/CAD), financial modeling (4.2x K3). Highspeed version 2-3x faster, 2x price/quota cost. Mandatory thinking mode (low/high/max). Open weights Nov 4.

•

7:17 — Pricing Analysis: Token Cost vs Task Cost: API: $2/$10 per 1M in/out (33% cheaper than K3, half Opus 5.5, 1/5 GPT-6 Astra). But domestic rivals 8-16x cheaper (DeepSeek Flash 0.02 RMB cached). Artificial Analysis: avg task cost $1.98 (same as K3) due to 14% more output tokens from longer reasoning.

•

10:53 — Compute Constraints: The Structural Bottleneck: Root cause: US chip sanctions since 2022 limit advanced GPU access. 2024 anecdote: Moonshot co-founder turned down enterprise deal — 'compute not enough.' China Token daily consumption 4x YoY (500T), normalized price doubled. Domestic chips (Huawei Ascend 950DT, Cambricon, T-Head) advancing but demand exceeds supply (orders to 2030). DeepSeek planning 160k Ascend 950DT; Zhipu using 100k+ domestic GPUs for self-optimizing inference.

•

15:53 — Strategic Dilemma & Conclusion: Kimi bets on high-cost/high-capability like OpenAI/Anthropic, but squeezed between cheaper domestic models and superior foreign flagships. Flash model in development; next flagship to scale params. Core question: allocate scarce compute to stronger model or affordable inference? Advice: test with own tasks before committing.

Speakers

•

大飞 (Da Fei) — Host / Analyst

⋮