China’s AI ecosystem stopped being predictable this week. Ant Group’s InclusionAI open-sourced a trillion-parameter reasoning model under the MIT License โ€” an unprecedented move from a payments infrastructure team. DeepSeek laid out plans for its first multimodal model with enterprise toolchain support, and Presidents Trump and Xi struck a deal reversing Biden-era chip restrictions, allowing Nvidia H200 shipments back into China with a 25% tariff. Here is this week’s China AI Weekly.

Ring-2.6-1T: Ant Group’s Trillion-Parameter Bombshell

The biggest story came from an unexpected source. InclusionAI, the technology arm of Ant Group (parent of Alipay), released Ring-2.6-1T โ€” a trillion-parameter Mixture-of-Experts reasoning model under the MIT License. This is the most permissive open-weight release at this scale in the industry’s history.

Key specs:

  • ~1 trillion total parameters, ~63 billion active per token, keeping inference costs practical despite the parameter count
  • 128K native context (256K via YaRN extension)
  • Adaptive reasoning-effort modes: “High” and “xHigh” for dynamic compute allocation
  • Released on Hugging Face under inclusionAI/Ring-2.6-1T with no usage restrictions

Vendor-reported benchmarks show Ring-2.6-1T competitive with DeepSeek-V4-Pro Max, Kimi K2.6 Thinking, and GPT-5.4 on AIME 2026, with notably stronger scores on ARC-AGI-2 and SWE-bench. (Source: Gigazine) (Source: Codersera)

Why This Matters

A trillion-parameter MIT-licensed model from a team most of the industry hadn’t noticed fundamentally changes the open-weight calculus. The MIT License means enterprises can deploy and fine-tune without usage restrictions โ€” a significant advantage for regulated industries like financial services and healthcare. Ant Group’s financial infrastructure DNA (billions of Alipay transactions annually) also brings production-grade reliability experience to the model’s development. For broader context on how Chinese AI is reshaping enterprise strategy, see our China AI ecosystem overview.

Ring-2.6-1T adds a credible fourth option to the open-weight frontier conversation alongside DeepSeek, Qwen, and Kimi โ€” and it arrived from a team no one was watching. The obvious caveat: these are vendor-reported benchmarks. Independent verification from LMSys Chatbot Arena will determine whether the scores hold up under standardized testing.

DeepSeek V4.1: Multimodal and MCP in June

DeepSeek confirmed it will ship V4.1 in June, adding native image and audio processing โ€” a first for the DeepSeek line and a significant evolution from the text-only V3 and R1 models. The update also brings Model Context Protocol (MCP) support and an expanded enterprise toolchain, marking DeepSeek’s shift from research demo to commercial product. For enterprise teams evaluating DeepSeek against Western alternatives, see our DeepSeek V4 CIO decision framework. The accelerated cadence follows a 140-day gap during which competitors shipped roughly 50 new models. (Source: aibase.com)

Pricing extension. DeepSeek extended its 75% discount on V4-Pro through May 31, 2026. Cache-hit pricing dropped to 1/10 of launch rates ($0.0028/1M tokens). Current rates are $0.435/1M input and $0.87/1M output tokens โ€” roughly 70% below comparable Western models. (Source: DeepSeek API Docs)

Funding picture. Founder Liang Wenfeng personally invested 20B yuan ($2.8B) into the company. DeepSeek is reportedly seeking 50B yuan ($7B) in its next round โ€” the largest single-round raise by a Chinese AI startup โ€” with the state-backed National Integrated Circuit Industry Investment Fund, Tencent, and Alibaba in discussions. The Big Fund’s involvement ties DeepSeek’s trajectory to Beijing’s domestic semiconductor ambitions, particularly since V4 was trained on Huawei Ascend hardware.

For enterprise teams, V4.1 is DeepSeek’s first enterprise positioning play. If it delivers on multimodal quality and MCP integration, it becomes a direct competitor to GPT and Claude for teams that prioritize API consumption over self-hosted deployment. The aggressive pricing trajectory also suggests inference costs will continue falling.

Trump-Xi Chip Deal: H200 Exports Resume with a Tariff

In a May 14 Beijing meeting, Presidents Trump and Xi agreed to allow Nvidia to export H200 chips to “approved customers in China” with a 25% tariff paid to the US โ€” the most significant US-China chip policy reversal since the Biden-era restrictions. Democrats in Congress criticized the move as a national security risk, setting up potential legislative challenges. (Source: Economic Times)

Three implications:

  • Immediate capacity relief. H200 access supplements domestic production as Huawei ramps toward 600,000 Ascend 910C units in 2026 and SMIC runs at capacity on mature nodes. Demand from labs like DeepSeek, Moonshot, and InclusionAI still outstrips domestic supply, and H200s fill a gap domestic fabs cannot yet close.

  • The 25% tariff is a policy experiment. It creates direct US Treasury revenue on advanced chip exports โ€” a mechanism with no precedent that could be adjusted, expanded, or revoked. Implementation details (who qualifies as “approved,” how tariffs are collected) remain unspecified.

  • Geopolitical whiplash risk. The reversal from full ban to tariff-based access complicates long-term capacity planning. Congressional opposition means the deal could face legislative challenges within months. For teams building on Chinese AI infrastructure, maintaining domestic chip compatibility remains a prudent hedge.

The underlying supply chain dynamics reinforce this: Huawei plans to double 2025’s top-end chip output, with total Ascend line production reaching up to 1.6 million dies in 2026. DeepSeek V4’s Day-0 compatibility across eight Chinese chipmakers demonstrates that domestic alternatives are viable, even if supply-constrained.

Open Source & Community

  • Ring-2.6-1T is the week’s standout open-weight release โ€” a trillion-parameter MIT-licensed model from outside the usual frontier labs. Third-party verification will determine whether it holds top-tier rankings, but the release alone resets expectations for what open-weight models can achieve.
  • DeepSeek models now exceed 75 million Hugging Face downloads, maintaining the top spot for Chinese AI families on the platform. The R2 32B model’s ability to run on 24GB consumer GPUs keeps it a go-to for cost-sensitive reasoning workloads.
  • Chinese labs collectively challenge the “frontier means closed” thesis. Between DeepSeek’s R2 and V4, Qwen3.6, Kimi K2.5, and now Ring-2.6-1T, the volume and capability of open Chinese models continues accelerating. Each release pushes the open-weight ceiling higher, making it harder for Western closed-source vendors to justify the premium without proportional capability advantages.

What to Watch

  • DeepSeek V4.1 delivery (June). If multimodal and MCP features land well, this is DeepSeek’s first enterprise-positioned model. Watch for third-party benchmarks on vision and audio tasks.
  • Ring-2.6-1T independent benchmarks. LMSys Chatbot Arena and Artificial Analysis results will determine whether vendor-reported scores against GPT-5.4 hold up.
  • Chip deal implementation. The “approved customer” framework and 25% tariff need published specifics. Watch for Commerce Department guidance and Congressional hearings.
  • Huawei Ascend production. With 600,000 910C units planned and hyperscaler orders placed, the second half of 2026 will test whether domestic supply can meaningfully close the gap with resumed Nvidia access.

That is this week’s China AI Weekly โ€” a fintech team dropped a trillion-parameter MIT-licensed model, DeepSeek confirmed its multimodal roadmap, and the US-China chip landscape flipped on its head. The common thread is accelerating pace: open-weight AI boundaries keep expanding, and the infrastructure race โ€” chips, models, and deployment pathways โ€” shows no signs of slowing.

Evaluating Chinese AI models for your enterprise stack? Big Hat Group delivers enterprise AI consulting with vendor-neutral orchestration, Azure-native deployments, and documented governance. Whether you are benchmarking Ring-2.6-1T, integrating DeepSeek V4.1, or building a multi-model strategy that navigates the shifting chip policy landscape, book a discovery call. Thirty minutes. No pitch deck. Just your use case and our bench.

Sources: gigazine.net, codersera.com (Ring-2.6-1T), aibase.com (DeepSeek V4.1), deepseek.com/api-docs (pricing), Economic Times (Trump-Xi chip deal), Bloomberg (Huawei Ascend), SCMP (SMIC), Hugging Face (model statistics)