TL;DR – What happened after the DeepSeek Moment?

One year after DeepSeek R1’s open‑source release, China’s AI community shifted from chasing the biggest single‑model performance to building flexible, cost‑effective, and hardware‑aware AI systems.

• Mixture‑of‑Experts (MoE) became the default architecture – it lets huge models stay affordable by activating only a subset of “experts” per request.
• Multimodal races exploded – text‑to‑image, video, audio, 3‑D and agent models were released in parallel, each bundled with full toolchains.
• Small models (≤30 B) surged – they are easy to run locally, fine‑tune, and embed in business workflows; large MoE models act as “teacher” nets for distillation.
• Apache 2.0 / MIT licenses now dominate – permissive terms removed legal friction, accelerating commercial adoption.
• Hardware‑first mindset – releases now ship with quantization, inference, and serving stacks tuned for domestic chips (Huawei Ascend, Cambricon, Kunlun, etc.), and training pipelines are openly documented.

These trends indicate that the competitive edge is moving from raw model size to system design, deployment efficiency, and open‑source ecosystem integration.

1. MoE – The Architecture of Choice

"Strong reasoning could be open, reproducible, and engineered in practice." – DeepSeek R1 team

• Why MoE? It partitions a gigantic parameter pool into many experts that are sparsely activated. A single inference only touches a fraction of the total compute, cutting inference cost dramatically while preserving the capability of a 100 B‑plus model.
• Real‑world impact: Models such as Kimi K2, MiniMax M2, and Qwen‑3 all adopted MoE, enabling:
  • Dynamic scaling of compute per request (simple queries vs. complex reasoning).
  • Deployment on heterogeneous hardware – cheap edge devices can run the same model by using fewer experts.
• Strategic outcome: China prioritized sustainable performance over absolute benchmark supremacy, aligning with limited compute budgets and export‑control pressures.

2. The Modality Arms Race

"The parallel emergence of video generation tools, 3‑D components, distillation datasets, and agent frameworks pointed to reusable system‑level capabilities." – Blog author

Modality	Notable Open Releases (2025‑2026)	Key Features
Text‑to‑Image / Image‑to‑Video	Step‑1X‑Edit, Hunyuan Video	High‑fidelity generation, editing APIs, edge‑to‑cloud pipelines
Audio / Speech‑to‑Speech	Step‑Audio‑R1.1, Step‑Audio‑R1	State‑of‑the‑art TTS & voice conversion, open evaluation suites
3‑D & Virtual Worlds	Hunyuan 3D, GLM‑Image (vision‑augmented)	Mesh generation, texture synthesis, real‑time rendering support
Agents / Tool‑Use	Various open‑source agent frameworks (e.g., LangChain‑China)	Integrated tool‑calling, memory, and planning modules

The community stopped publishing just weights. Every release bundled:

• Inference scripts for multiple runtimes (ONNX, TensorRT, FastDeploy).
• Dataset pipelines for fine‑tuning across modalities.
• Evaluation harnesses (e.g., multimodal benchmarks, human‑in‑the‑loop tests).

3. Small Models – The Workhorse Tier

"In environments with limited compute or strict compliance requirements, these models were far better suited for long‑term operation."

• Parameter range: 0.5 B – 30 B.
• Why they matter:
  • Easy to run on a single GPU or even on high‑end CPUs.
  • Faster fine‑tuning → quicker product iteration.
  • Lower compliance risk (smaller attack surface, easier auditing).
• Ecosystem pattern: Large MoE models (100 B – 700 B) serve as teacher nets. Knowledge is distilled into dozens of compact models, creating a pyramid:
  • Top: Massive MoE models – research showcase, capability ceiling.
  • Middle: 30 B‑70 B distilled models – high‑quality APIs for SaaS.
  • Base: ≤30 B models – on‑premise, edge, and embedded applications.
• Community metric: The share of small‑model downloads in the monthly Hugging Face summary grew from ~35 % (early 2025) to >55 % (mid 2026).

4. License Liberalisation – Apache 2.0 Takes Over

"More permissive licenses lowered the friction around using, modifying, and deploying models in production."

• Before DeepSeek R1: A mix of custom, restrictive licenses (e.g., CC‑BY‑NC‑SA). Legal uncertainty slowed corporate adoption.
• Afterward: Apache 2.0 and MIT became the de‑facto defaults for Chinese open‑source AI models.
• Effect:
  • Companies can embed models in proprietary products without relicensing.
  • Cloud providers can offer “model‑as‑a‑service” on domestic infra without negotiating per‑model terms.
  • Academic collaborations across borders become smoother (license compatibility with GPL, BSD, etc.).
• Data point: In the “Open‑Source Heatmap” the proportion of Apache‑licensed releases jumped from 12 % (Jan 2025) to 68 % (Oct 2026).

5. From Model‑First to Hardware‑First

"A prominent goal was no longer just to make weights downloadable, but to ensure that models could run directly on target domestic hardware."

5.1 Inference‑First Bundles

• Day‑Zero hardware support – DeepSeek‑V3.2‑Exp shipped with ready‑to‑run binaries for Huawei Ascend and Cambricon chips.
• Quantization pipelines – Models released with INT4/INT8 versions tuned for specific ASICs, dramatically reducing latency and power.
• Serving stacks:
  • Mooncake (Moonshot AI) – open‑source pre‑fill/decoding separation, GPU‑/CPU‑aware scheduling.
  • FastDeploy 2.0 (Baidu) – extreme quantization + cluster‑level optimisations.
  • Qwen ecosystem (Alibaba) – end‑to‑end CI/CD from model repo → Docker image → cloud deployment.

5.2 Training‑Side Transparency

• Ant Group – Ling: Documented a 5 % cost reduction per trillion tokens using domestic AI chips.
• Baidu – Qianfan‑VL: Published the full topology of a 5 000‑node Kunlun P800 cluster, including pipeline parallelism strategies.
• Zhipu – GLM‑Image & Telecom – TeleChat3: Both announced entirely domestic‑chip training, signaling that China’s chip ecosystem now supports the full training stack for trillion‑token models.

5.3 Strategic Implications

• Resilience to export controls: By co‑optimising models and chips, Chinese firms reduce dependence on NVIDIA/H100‑class GPUs.
• Ecosystem lock‑in: Open‑source serving tools (Mooncake, FastDeploy) are tightly coupled with domestic hardware SDKs, creating a virtuous cycle of adoption.
• Remaining bottlenecks: Reports of compute scarcity (e.g., Zhipu limiting sign‑ups) show that hardware supply still lags behind demand, a risk for scaling.

6. Reconstruction in Progress – The New Competitive Landscape

"Chinese companies are no longer optimizing isolated models. Instead, they are pursuing distinct architectural paths aimed at building full ecosystems suited to an open‑source world."

• System‑level competition has overtaken raw benchmark races. Success is measured by:
  1. Deployment cost per query (USD/1 M tokens).
  2. Time‑to‑market for fine‑tuned variants (days vs. weeks).
  3. Hardware‑agnostic portability (single repo works on Ascend, Cambricon, and NVIDIA).
• Open‑source as strategic asset: By publishing not just weights but the entire stack (datasets, training scripts, inference engines), firms create network effects that attract developers, startups, and academia.
• Future outlook (2026+):
  • Hybrid MoE‑Quantized models that dynamically switch precision per expert.
  • Edge‑centric agents running on low‑power ASICs for real‑time decision making.
  • Cross‑modality distillation pipelines that compress a multimodal teacher into a unified small model.

---

Takeaways for Researchers, Developers, and Policymakers

1. Focus on system design – When contributing to open‑source AI, prioritize reproducible pipelines, hardware‑aware quantization, and serving scripts.
2. Leverage small models – They dominate real‑world deployments; distillation from large MoE teachers is a proven path to high quality.
3. Watch licensing – Apache 2.0 and MIT will remain the safest choices for commercial integration.
4. Monitor hardware policy – Export‑control shifts will continue to reshape where and how large‑scale training occurs.
5. Collaborate across modalities – Multimodal open‑source projects are converging; cross‑modality datasets and evaluation suites are becoming community standards.

The DeepSeek Moment sparked a re‑architecture of China’s AI ecosystem: from isolated, monolithic models to a layered, hardware‑first, open‑source stack that balances capability, cost, and strategic independence.