Region Focus | China

🌏 中文版

Region: China

The keyword for China's AI Agent ecosystem this week is "opening up and tightening down at the same time." DeepSeek open-sourced its Agent execution infrastructure while sharply raising model API prices. Zhipu's new model turned out to be so good at finding vulnerabilities that it delayed its planned open-source release. And Beijing loosened chip import controls just a crack. These three stories may seem contradictory, but they point to the same signal -- China's AI industry is shifting from "competing on benchmark scores" to "competing on who controls the most chips across the entire value chain."

Key Developments

DeepSeek Open-Sources Agent Execution Framework DeepSeek Harness; API Prices Rise Up to 1,100% the Same Week

On the evening of August 13 (Beijing time), DeepSeek released the developer preview of DeepSeek Harness (DSH) under the MIT license. This is not model weights -- it is the full Agent runtime infrastructure surrounding the model: model, tools, skills, sessions, sandbox, storage, Agent loop, and UI are all swappable plugins built on DeepSeek's in-house Cordis plugin system, with four preset modes: Standard, Minimal (PTC -- Programmatic Tool Calling), and Creative. DeepSeek even launched a separate WeChat account ("DeepSeek Harness Team") with a black whale logo to differentiate from the blue whale used for DeepSeek models, signaling the company's intent to build Harness into an independent developer ecosystem. The launch was explosive: GitHub stars surpassed 20,000 within one hour (the fastest ever), accumulating roughly 158,000 stars by week's end. The community submitted 2,000+ plugin proposals in two days, and approximately 300 plugins were developed during the brief beta period. (36Kr -- Harness architecture, 36Kr -- hands-on review, MarkTechPost)

Almost simultaneously, DeepSeek announced that V4-Pro/V4-Flash API prices would rise across the board starting midnight Beijing time on August 17, introducing peak/off-peak pricing for the first time (peak: 9:00--12:00 and 14:00--18:00 Beijing time; off-peak rates are half of peak). For V4-Pro, peak cached-input pricing jumped from RMB 0.025 to RMB 0.3 per million tokens (a 1,100% increase), and output pricing rose from RMB 6 to RMB 27 (up 350%). DeepSeek's official explanation was "more rational allocation of computing resources," but multiple Chinese outlets noted this is tied to ongoing domestic compute shortages -- and it is not an isolated event: Zhipu has raised prices three times this year, and Moonshot's Kimi K3 launch came with input prices tripling and output prices nearly quadrupling. This pricing wave marks a shift in China's LLM competition from "price war" back to "value war." (PTT/Business Media, Yahoo Finance Taiwan, Anue)

Alibaba Open-Sources Flagship Weights Across the Board; Qwen3.8 Max Tops Long-Context Leaderboard

Alibaba Cloud shipped a wave of releases this week: Qwen3.8 Max topped the LongBench v2 long-context reasoning leaderboard at 66.3%, surpassing Claude Opus 4.5 (64.4%) and the previous-generation Qwen3.5 397B (63.2%), while open-sourcing the flagship model's weights. A second model, Qwen3.8-27B, ships under Apache 2.0 with vision support and runs on 17 GB of VRAM (i.e., laptop-grade hardware), scoring 52 on the Artificial Analysis Intelligence Index -- seen as a major milestone for mid-size open models. Alibaba's Qianwen Office team also open-sourced MyContext, a context infrastructure system that runs locally on the user's own device, automatically organizing IM conversations, documents, and collaboration records into retrievable, continuously updated working files, aiming to reduce the hallucination and contradictory-information problems common in long-running Agent tasks. (CNBC, Artificial Analysis)

Zhipu GLM-5.3's Cybersecurity Capabilities 'Emerge Unexpectedly'; Open-Source Plans Delayed

On August 14, Zhipu released GLM-5.3, which shares the same base model as GLM-5.2 -- all capability improvements come from scaled post-training. Coding performance rose roughly 50% over the previous generation, with DeepSWE climbing from 46.2% to 66.9% and Terminal-Bench 3.0 jumping from 4.6% to 28.3%. What caused Zhipu to hit the brakes was what the company called "unexpectedly emergent" cybersecurity capabilities: CyberGym vulnerability discovery scored 84.5% (first among open models, up from 77.2% for 5.2), surpassing even Zhipu's own Anthropic reference model benchmark (83.8%); ExploitBench vulnerability exploitation doubled from 24.4% (5.2) to 54.4%. Zhipu disclosed that the GLM series has scanned 269 open-source projects and discovered 2,436 vulnerabilities, of which 1,097 are rated high or critical severity. As a result, Zhipu postponed the planned weight release by approximately two weeks, citing the need to assess the cybersecurity risks of open-weighting -- the first time a top Chinese model lab has delayed open-sourcing because a model was "too good at attacking." (36Kr -- hands-on review, 36Kr -- delayed open-source story)

Markets reacted positively: in an August 17 report, Goldman Sachs called GLM-5.3 "yet another major leap in Chinese AI model progress," maintaining a Neutral rating on Zhipu with a target price of HKD 1,610. A clarification is warranted: some Chinese-language reports circulating this week claimed Zhipu's market cap "exceeded RMB 1 trillion," but cross-referencing shows this conflates two events. Zhipu's market cap crossed the 1-trillion threshold in June (driven by GLM-5.2), and the unit was HKD, not RMB. Goldman Sachs' August 17 report pegs the latest market cap at approximately USD 75 billion -- above DeepSeek's latest valuation (USD 50 billion) and MiniMax (USD 14 billion) but, when converted, below the "RMB 1 trillion" scale. (Yahoo Finance HK, Caixin)

ByteDance, Tencent Approved to Import H200; Chip Controls Ease at the Margin

The Financial Times reported on August 19 that the Chinese government has approved a limited import of NVIDIA H200 chips, with ByteDance and Tencent each receiving approximately 10,000 units in recent weeks; other Chinese tech companies may follow soon. Notably, the U.S. had originally approved up to 100,000 H200s per Chinese company, but Beijing wants most chips to remain outside mainland China (e.g., in Hong Kong) to support domestic chipmakers such as Huawei. The roughly 10,000 chips actually delivered to ByteDance and Tencent represent only about 10% of the U.S.-approved ceiling. Purchases still require approval from the National Development and Reform Commission (NDRC). Sources cited in the report noted that Chinese labs increasingly rely on domestic chips for inference, but training frontier models still heavily depends on NVIDIA architecture -- which explains why Beijing is willing to loosen controls now: models recently released by Moonshot K3, Alibaba, DeepSeek, and Z.ai (Zhipu) show shrinking capability gaps, and Beijing does not want chip supply to become a bottleneck holding back top labs. (HK01, World Journal)

Deep Analysis

The most noteworthy signal from China's AI Agent ecosystem this week is that "openness" is becoming a carefully calculated strategic lever rather than a pure technical ideal. Applying Porter's Five Forces to this week's four stories reveals a consistent logic:

Threat of new entrants (entry barriers): DeepSeek's open-sourcing of Harness ostensibly lowers the barrier to "building a usable Agent product," but the real barrier has shifted to the plugin ecosystem -- with 2,000+ plugin proposals flooding in within two days, those who stake out ecosystem positions first build network effects that latecomers cannot replicate. This is fundamentally different from the competitive logic of open-weight models (anyone who downloads can use them).

Supplier bargaining power: The H200 import relaxation is not "chip liberalization" -- it is a bargaining lever Beijing actively modulates. By deliberately keeping most approved chips in Hong Kong rather than the mainland, Beijing simultaneously maintains purchasing dependence on NVIDIA and protects bargaining space for domestic suppliers like Huawei, turning chips into a policy tool it can tighten or loosen at will.

Threat of substitutes: Zhipu's delay of GLM-5.3 open-sourcing is fundamentally about managing the speed at which "open-weight models" substitute for "controlled closed-source services." Once weights are public, anyone can strip safety guardrails and attach attack tooling -- substitution would happen faster than Zhipu could assess the risks, so they chose to slow down first.

Competitive rivalry: DeepSeek raising API prices in the same week it open-sourced Harness seems contradictory but actually splits the competitive arena in two -- using a free, open-source execution layer to capture developer mindshare (not optimizing for near-term revenue) while using a repriced model layer to harvest pricing power (model capabilities now justify paid access). This "split the stack, fight on two fronts" approach is harder for competitors to replicate than either price hikes or open-sourcing alone.

Takeaways for Taiwan Founders

• If you are building an Agent framework or developer tool: DeepSeek Harness's play of "trading an open-source execution layer for developer mindshare" is worth studying, but note that the Chinese playbook assumes you also own a cheap base model you can monetize. Taiwan teams without in-house models are better off focusing on the plugin ecosystem or vertical integration layer rather than rebuilding a Harness.
• If your product relies on Chinese open-source models like DeepSeek, Qwen, or GLM for cost optimization: this pricing wave (DeepSeek peak rates up 1,100%, Zhipu raising prices three times this year) means the assumption that "Chinese open-source models will always be cheaper than U.S. alternatives" no longer holds. Cost models need to factor in peak/off-peak pricing, exchange rate risk, and supplier concentration risk. Do not treat a single vendor's low prices as a long-term moat.
• If you are building cybersecurity or vulnerability scanning products: GLM-5.3's "emergent" cybersecurity capabilities are a signal -- general-purpose coding models are rapidly approaching or even surpassing dedicated security models in vulnerability discovery. Taiwan security startups still using "AI-assisted" as a differentiator may see their moat eroded by general-purpose LLMs faster than expected. The defensible position lies in moving toward "judgment and remediation workflows above the model."

Takeaway

I used to think Chinese model companies' "open-source" strategy was simply an extension of their low-price approach -- giving things away for free to grab market share. Looking at this week's developments, I now see that open-sourcing has been decomposed into more finely calibrated levers: DeepSeek trades open-source Harness for developer mindshare while raising model prices to harvest revenue; Zhipu does the opposite, using "delayed open-sourcing" to manage risk and brand trust. Whether to open-source or not is no longer an ideological choice in China -- it is a pricing and risk-management strategy each company calculates based on its position in the value chain.

References

• 36Kr -- Why is DeepSeek's Harness a black whale?
• 36Kr -- DeepSeek Harness is here, and it doesn't want to be the next Codex
• MarkTechPost -- DeepSeek releases DeepSeek Harness developer preview
• PTT/Business Media -- DeepSeek model API price hike exceeds 1,100%
• Yahoo Finance Taiwan -- DeepSeek V4 API price hike next week
• Anue -- DeepSeek price increase takes effect, up to 1,100%
• CNBC -- Alibaba open-sources Qwen open-weight AI laptop models
• Artificial Analysis -- Qwen3.8-27B
• 36Kr -- Hands-on GLM-5.3: same model, why does it feel like two AIs
• 36Kr -- Because the new AI version is too powerful, Zhipu temporarily doesn't dare open-source it
• Yahoo Finance HK -- Goldman Sachs: GLM-5.3 is yet another major leap in Chinese AI model progress
• Caixin -- Zhipu's market cap exceeds 1 trillion: does the 'China's Anthropic' valuation hold up?
• HK01 -- FT: China allows limited Nvidia H200 chip imports, Tencent and ByteDance each get 10,000
• World Journal -- FT: China eases Nvidia H200 chip controls, ByteDance and Tencent each get 10,000