This was the week China’s AI ecosystem delivered two signal events: DeepSeek โ€” the famously self-funded, open-weight disruptor โ€” opened its doors to outside capital in what could become the largest private AI fundraise in history, while Alibaba flexed its platform muscle with a new multimodal agent model and a strategic opening of the Qwen app to external developers. Meanwhile, MiniMax and StepFun shipped significant open-weight releases that continue to raise the capability ceiling for Chinese open-source AI. This is the China AI Weekly for June 6, 2026.

DeepSeek Opens the Kimono: $7โ€“10B+ First External Round

The biggest story in Chinese AI this week โ€” and arguably the biggest story in global AI infrastructure โ€” is DeepSeek initiating its first-ever external funding round. Reports peg the raise at approximately $7โ€“10 billion (50โ€“70 billion yuan) at a valuation between $45โ€“59 billion, making it potentially the largest single financing round in Chinese tech startup history. (Bloomberg, 36Kr, Outlook Business)

The investor lineup is as strategic as it gets: Tencent is reportedly committing ~$1.4 billion (100 billion yuan), CATL (Contemporary Amperex Technology) is in for ~$700 million, and the National AI Industry Investment Fund (Big Fund) is anchoring the round. Founder Liang Wenfeng is personally injecting an additional ~$2.8 billion to retain control, signaling that capital comes with strings โ€” but not control. (163.com, Crossing the River)

Why this matters: DeepSeek has been entirely self-funded through Liang’s hedge fund, Zhejiang High-Flyer, since its founding in 2023. The decision to take outside capital is a structural pivot โ€” and the commitments Liang reportedly made to investors are worth reading closely:

  • Open-weight commitment: Liang pledged to continue shipping open-weight models. No pivot to closed-source.
  • AGI priority: The company will prioritize research and frontier capability over aggressive commercialization โ€” no “build enterprise sales in 18 months and IPO” playbook.
  • Price floor is now strategic: The permanent 75% V4 Pro price cut โ€” announced just days before this round โ€” was not a promotion. DeepSeek now has the capital to sustain the lowest frontier inference pricing indefinitely. At $0.87 per million output tokens for V4 Pro, the pricing baseline for agent-tier inference just got a multi-year anchor.

What to watch: Tencent’s involvement is particularly significant. Tencent has been quietly building its own AI stack while investing across the ecosystem. A strategic position in DeepSeek gives Tencent access to frontier models without needing to win the talent war for its own foundation model lab โ€” while simultaneously integrating agentic AI into WeChat, the 1.3-billion-user super-app that Tencent is actively retrofitting with AI agent capabilities. (Nikkei Asia)

Alibaba: Qwen3.7-Plus Launches, Qwen App Opens to External Developers

Alibaba had a multi-front week. On June 2, the Qwen team released Qwen3.7-Plus, the multimodal sibling to the text-only Qwen3.7-Max that launched in May. Qwen3.7-Plus adds image and video understanding, deep reasoning, self-programming, tool invocation, verification, and autonomous iteration โ€” all accessible through the Bailian platform (Alibaba Cloud’s Model Studio for international users). (MarkTechPost, Qwen Blog)

The model’s preview ranked #16 overall in Vision Arena (LM Arena), placing Alibaba as the #5 lab in vision. For enterprise workloads involving OCR, chart reading, or video-frame analysis, Qwen3.7-Plus represents the most capable Chinese multimodal API available at $2.50 per million input tokens โ€” roughly half the cost of comparable Western frontier tiers. (TechFastForward)

Then on June 3, Alibaba announced it would open the Qwen consumer app to external agents and “skills” โ€” a direct challenge to Tencent’s WeChat agent strategy. Third-party developers can now build and publish agent capabilities that run inside Qwen, creating an app-store-like ecosystem for agent skills. This is Alibaba’s bid to own the consumer AI agent platform layer, parallel to its enterprise agent play on Bailian. (Nikkei Asia, Channel Post MEA)

On June 6, Alibaba pitched Qwen3.7-Plus as a “computer-use AI agent” โ€” positioning it for automated desktop and browser-based workflows in the same category as Anthropic’s computer-use capabilities. (WinBuzzer)

Why this matters for enterprise teams: Alibaba is executing a platform play, not just a model play. The combination of competitive pricing ($2.50/$7.50 per million tokens for Max), multimodal agent capabilities, external developer ecosystem, and desktop automation positioning makes Qwen the most complete Chinese AI platform for teams evaluating multi-model strategies. The 35-hour autonomous runtime demonstrated for Qwen3.7-Max on complex coding tasks (VentureBeat) is a specific capability that enterprises running long-horizon automation should benchmark directly.

Open Source & Community: MiniMax M3, Step-3.7-Flash, and the Open-Weight Wave

The open-weight release cadence from Chinese labs continued at full speed this week with two standout launches:

MiniMax M3 โ€” 1M Context, Native Multimodal, Open-Weight

MiniMax released M3, a model that checks every box: 1-million-token context window via the novel MSA (MiniMax Sparse Attention) architecture, native multimodal understanding (images + video), and computer-use capabilities. On SWE-Bench Pro, M3 scored 59.0% โ€” beating GPT-5.5 and Gemini 3.1 Pro, and approaching Claude Opus 4.7. On Terminal-Bench 2.1, it scored 66.0%. (MiniMax)

The model can operate autonomously for extended periods โ€” demonstrated by independently reproducing an ICLR 2025 Outstanding Paper (12 hours, 18 commits, 23 experimental charts) and optimizing an FP8 GEMM CUDA kernel on NVIDIA Hopper hardware over 24 hours across 1,959 tool calls.

Why this matters: M3 is the first Chinese model to simultaneously deliver frontier coding benchmarks, 1M context, native multimodal, and open weights. For teams self-hosting on NVIDIA hardware, M3 is now a serious alternative to Western open-weight models.

StepFun Step-3.7-Flash โ€” Open-Weight Agent Model at 400 Tokens/s

StepFun (Shanghai-based, one of the best-funded emerging labs) open-sourced Step-3.7-Flash, a 198-billion-parameter sparse MoE model with only 11 billion active parameters per inference step. Peak generation speed hits 400 tokens per second โ€” roughly 2ร— Gemini 3.5 Flash and 4ร— most mainstream models. Context window: 256K tokens. (AI Puzi)

The model is specifically optimized for production-grade agent workloads: reliable long-chain tool calling, multimodal input (images, video, UI screenshots), and compatibility with mainstream agent frameworks including Claude Code, KiloCode, RooCode, OpenCode, and OpenClaw. It also supports MCP protocol and Skills-based tool invocation.

Why this matters: At 400 tokens/s with 11B active parameters, Step-3.7-Flash changes the cost equation for agent deployments at scale. StepFun claims 97% of Claude Opus 4.6 coding performance at 1/9th the cost on SWE-Bench. For teams running high-volume agent inference, that’s a number worth stress-testing.

Also This Week

  • Alibaba open-sourced Qwen-VLA โ€” a Vision-Language-Action model for robotics, unifying perception, understanding, and action in a single model. (AI Puzi)
  • Ant Group (่š‚่š็™พ็ต) open-sourced Ling-2.6-1T โ€” a trillion-parameter model achieving SOTA on AIME 2026 and SWE-Bench, optimized for agent and coding tasks. (Prompt Yuzhou)
  • Baidu released PaddleOCR-VL-1.6 โ€” document parsing model with 96.33% accuracy, setting a new SOTA. (AI Base)

What to Watch

  • DeepSeek funding finalization โ€” The round is reported but not yet closed or confirmed by DeepSeek. Watch for formal confirmation, which would set a definitive valuation and reveal the full investor syndicate. The $45โ€“59B valuation range is wide โ€” the final number matters.
  • Tencent’s WeChat agent strategy โ€” With Tencent now double-invested (WeChat agent integration + DeepSeek stake), expect concrete WeChat AI agent features to surface in the coming weeks. This is the most consequential consumer AI distribution battle in Asia.
  • Qwen app ecosystem โ€” The external skills/store strategy for Qwen is a direct copy of how WeChat grew its mini-program ecosystem. Whether developer adoption materializes will determine whether Alibaba can create a consumer AI platform, not just an app.
  • Open-weight cluster effect โ€” With MiniMax M3, Step-3.7-Flash, and Ling-2.6-1T all landing within days of each other, teams evaluating self-hosted AI should run these against their production workloads now. The open-weight landscape has shifted meaningfully in the past two weeks.

From DeepSeek’s billion-dollar coming-out party to Alibaba’s three-pronged agent strategy to a wave of genuinely competitive open-weight releases, this was a week that made clear: Chinese AI labs are not just keeping pace โ€” they are setting the pace in specific areas (pricing, open-weight release cadence, agentic autonomy) that directly affect how enterprise teams outside China should plan their multi-model strategies.

Check back next week.