The Chinese AI Blitz: Qwen 3.5, GLM-5, and MiniMax Just Changed the Model Landscape

AI Models · February 11, 2026

The Chinese AI Blitz: Qwen 3.5, GLM-5, and MiniMax Just Changed the Model Landscape

Maya Chen  ·  9 min read

Why the Lunar New Year Timing Was Deliberate

Lunar New Year 2026 carried significant symbolic weight: the Year of the Horse, traditionally associated with speed and competitive drive. Chinese tech companies are acutely aware of how international observers read timing signals. Releasing three frontier-class models in 48 hours during a period of global focus on China — and specifically on Chinese AI — is not an accident of engineering schedules.

This mirrors the DeepSeek R1 moment from early 2025, when Chinese open-source capabilities blindsided Western analysts who had underestimated the pace of Chinese AI development. The Qwen / GLM-5 / MiniMax blitz is the next chapter of the same narrative: Chinese labs are not following the frontier, they are competing at it.

Qwen 3.5: Alibaba’s Multimodal Bet

Alibaba’s Qwen 3.5 flagship — Qwen3.5-397B-A17B — uses a Mixture of Experts architecture: 397 billion total parameters, with 17 billion active per inference pass. The MoE design means the model achieves large-parameter reasoning quality while keeping per-token compute costs manageable.

The strategic breakthrough is multimodality: Qwen 3.5 is the first Chinese model to natively handle text, image, and video input in a single architecture. Not bolted on, not a separate API endpoint — unified multimodal at scale. The smaller Qwen3.5-Medium model is benchmarking at parity with Anthropic Sonnet 4.5 on standard evaluations, which means Alibaba has built a genuinely competitive mid-tier offering, not just a flagship showcase.

GLM-5: The 754B Open-Source Bomb

Zhipu AI released GLM-5 on February 11 under an MIT license. At 754 billion parameters, it is the first Chinese open-weight model above the 700B threshold — and one of the largest MIT-licensed models ever released by any lab, anywhere. The MIT license is the significant detail: it means any organisation, anywhere, can run GLM-5 commercially without licensing fees or restrictions.

Zhipu’s framing for GLM-5 is deliberately pointed: it is designed for “agentic engineering”, not “vibe coding.” The distinction matters — it signals a model trained and evaluated for reliable multi-step tool use, code execution, and structured workflow completion, rather than impressive-sounding but unreliable code generation.

Context window: 200K tokens. Not the 1M of Claude Opus 4.6, but more than sufficient for most production agentic applications.

MiniMax: The Price Cutter That Performs

MiniMax released two models: M2.5 and the faster M2.5-Lightning. Both hit near-state-of-the-art benchmark performance at a price point that undercuts every major Western frontier model: $0.30 per million input tokens, $1.20 per million output tokens.

To contextualise: Claude Opus 4.6 input pricing is roughly 50x higher per token. GPT-5.2 is in a similar range. MiniMax is offering near-frontier performance at commodity pricing — which is a direct attack on the assumption that the best models will necessarily be the most expensive ones.

The post-release period brought controversy: on February 23, Anthropic alleged that MiniMax had distilled Claude — that is, trained M2.5 using outputs from Claude API calls in violation of Anthropic’s terms of service. MiniMax denied the allegation. No independent verification had been completed as of this writing.

Chinese Models vs Western Frontier: Comparison

What This Means for Western Labs

The Chinese AI blitz targets three specific Western advantages simultaneously:

For organisations evaluating LLM strategy: the competitive set is no longer “OpenAI vs Anthropic.” It includes Alibaba, Zhipu, MiniMax, and a pipeline of Chinese labs now operating at frontier scale. GLM-5 MIT especially is a landmark — any enterprise that needs a 700B+ parameter model can now deploy one commercially without a vendor relationship. That’s a structural shift in how the AI supply chain works.

Related Reading

→ Claude Opus 4.6 Review: Anthropic’s New Model Beats GPT-5.2
→ 2026 AI Arms Race Recap

The global AI race, tracked weekly