This is the first issue of China AI Watch — a weekly digest of what's actually happening inside China's AI ecosystem, compiled from Chinese-language primary sources before Western media picks it up. If you're building with Chinese AI models (or competing against them), this is your early warning system.
This Week at a Glance
| Event | Company | Impact | Urgency |
|---|---|---|---|
| Kimi K3 released — 2.8T open-weight model | Moonshot AI | Largest open-weight model ever shipped | 🔴 High |
| deepseek-chat / deepseek-reasoner legacy API deprecation | DeepSeek | Hard cutoff: July 24, 2026 (6 days left) | 🔴 High |
| GLM-5.2 fully open-sourced | Zhipu AI | Free commercial use, Apache 2.0 license | 🟡 Medium |
| MiniMax-M3 API now available | MiniMax | 428B MoE, 1M context, competitive coding | 🟡 Medium |
| Qwen3.7-Plus confirmed open-source, Max stays closed | Alibaba | Split strategy: open for mid-tier, closed for flagship | 🔵 Info |
1. Kimi K3: The 2.8 Trillion Parameter Elephant in the Room
On July 16, Moonshot AI (YueZhiAnMian) released Kimi K3 — the largest open-weight model anyone has ever shipped. The numbers are staggering: 2.8 trillion total parameters, a Mixture-of-Experts architecture with 896 experts (16 active per token), and a 1-million-token context window. Weights will be publicly available by July 27.
What the Chinese sources are saying
Chinese tech media is calling this Moonshot's "DeepSeek moment" (their DeepSeek moment). The comparison matters: DeepSeek's V3/R1 releases forced every US lab to slash prices. Moonshot is betting that an open-weight 2.8T model will do the same to the high end of the market.
Why this matters for builders
If you're running inference on Chinese models through APIs, K3's release puts downward pricing pressure on the entire ecosystem. DeepSeek, Zhipu, and Alibaba all have to respond within weeks, not months. Expect price cuts or capability bumps from competitors.
The catch
Reasoning is always on — you can't disable it. Moonshot's API only supports reasoning_effort: "max" today. For latency-sensitive applications (customer support bots, real-time agents), this is a dealbreaker. K3 is a coding/reasoning powerhouse, not a speed daemon.
2. DeepSeek Legacy API Deprecation: 6 Days to Migrate
Buried in DeepSeek's API changelog: the legacy model names deepseek-chat and deepseek-reasoner will be permanently discontinued on July 24, 2026. If you're still pointing to these model IDs, your integration breaks in 6 days.
Migration path
# Old (stops working July 24)
model = "deepseek-chat"
# New — pick one:
model = "deepseek-v4-flash" # Fast, cheap, 90% cache-hit rate
model = "deepseek-v4-pro" # Slower, stronger reasoning
model = "deepseek-r1-0528" # Reasoning-focusedWhy DeepSeek is doing this
From the Chinese developer community (V2EX, Juejin), the consensus is that DeepSeek is consolidating to reduce inference infrastructure costs. Supporting legacy model versions means keeping older checkpoint clusters running — expensive for a company competing on price. Killing legacy names forces everyone onto their newest (and most cost-efficient) serving infrastructure.
Action item
If you have any production code using deepseek-chat or deepseek-reasoner, migrate today. The new models are drop-in replacements with identical API schemas. No SDK changes needed — just update the model string.
3. GLM-5.2 Goes Full Open-Source
Zhipu AI (Zhipu) announced that GLM-5.2, their flagship model, is now fully open-source under Apache 2.0. This means free commercial use, modification, and redistribution. Weights are available on HuggingFace and ModelScope.
Context: Zhipu's survival play
Chinese tech media (TMTPost) published a piece this week titled "Zhipu's Last Stand" (ZhipuAGI). The company is burning cash and losing ground to DeepSeek's price aggression and Moonshot's capability leaps. Going full open-source is a community play: build ecosystem lock-in before running out of runway.
What you get
GLM-5.2 scores competitively on coding benchmarks. The GGUF quantized versions are already circulating — you can run a Q4 quant on a single RTX 4090. For teams wanting a self-hosted Chinese model without DeepSeek's MoE infrastructure requirements, GLM-5.2 is the most practical option.
4. MiniMax-M3: The Quiet Contender
MiniMax released M3 in June, and it's now available via first-party API and third-party providers (SiliconFlow, GMI, Novita). The specs: 428B parameters, MoE with ~23B active, 1M context window, and a custom attention mechanism called MiniMax Sparse Attention (MSA).
Benchmark positioning
M3 scores ~1670 on GDPval-AA — behind GPT-5.5 (1769) but level with Claude Sonnet 4.6. For coding tasks, it's competitive with DeepSeek V4 Pro. Weights haven't been released yet, but MiniMax says they're coming.
Why it's not getting more attention
Kimi K3's 2.8T announcement completely overshadowed M3 in both Chinese and English media. But for production workloads where you need a reliable API with 1M context and strong coding, M3 is arguably the better choice today — K3's weights aren't even public yet.
5. Qwen3.7: Alibaba's Split Strategy
Alibaba confirmed that Qwen3.7-Plus will ship as open-source, while Qwen3.7-Max (the flagship) stays proprietary. This mirrors what we're seeing across the Chinese AI ecosystem: open-source the mid-tier to build ecosystem adoption, monetize the frontier.
The pattern
Every major Chinese lab now follows the same playbook:
- Open-source: Mid-tier models (GLM-5.2, Qwen3.7-Plus, Kimi K3) to build developer mindshare
- Closed/API: Flagship models (Qwen3.7-Max, DeepSeek V4 Pro) for revenue
- Price war: Aggressive pricing on API access to undercut US providers
Alibaba's twist: they're the only player with the cloud infrastructure (Alibaba Cloud) to bundle model API + hosting + storage into one bill. Watch for deeper integration plays.
What This Means for You
If you're building products on Chinese AI APIs, this week crystallized three things:
- The price war is accelerating. K3's 2.8T open-weight release forces everyone to compete on value, not just price. Expect API cost cuts across the board within 30 days.
- Legacy APIs are dying. DeepSeek's 6-day deprecation notice is a signal: Chinese labs iterate fast and don't maintain backward compatibility. Build with model version abstraction from day one.
- Open-weight is the new normal. When a 2.8T model goes open-weight, the definition of "frontier" shifts. Self-hosting competitive models is now viable for teams with modest GPU budgets.
"The gap between what Chinese developers know and what Western developers know about Chinese AI is 3-7 days. That gap is your competitive advantage — if you close it." — China AI Watch, Issue #1
Next Week
We're watching: Kimi K3 weight release (July 27), DeepSeek's response to K3 (price cut or new model?), and Alibaba Cloud Summit announcements. If you're not subscribed, bookmark this page — new issues drop every Friday.