Three Major Chinese AI Models Incoming: Qwen 4, DeepSeek V4, and GLM Next Version Breakdown

Chinese AI Labs Accelerate: Three Flagship Models on the Horizon

The domestic AI race is heating up again. According to overseas AI news bloggers, Alibaba's Qwen, DeepSeek, and Zhipu (GLM) are all entering intensive testing and release windows for their next-generation models at nearly the same time. From mysterious anonymous models appearing on Code Arena to leaked screenshots of limited rollouts on social platforms, all signs point to the same conclusion: a concentrated wave of new Chinese AI model releases is imminent.

This article summarizes and analyzes information from related leak videos. It's worth noting that most of the following content remains unconfirmed rumors and third-party test inferences — readers should exercise their own judgment.

Qwen 3.8 and Qwen 4.0: Two Mystery Models Spotted in the Arena

Two anonymous test models have recently appeared on Code Arena — referred to as Caleb and Cydonian Alpha. Caleb is widely believed to be an early checkpoint of Alibaba's next-generation flagship Qwen model, with its performance on complex 3D code generation tasks described as "extremely impressive."

About Code Arena: Code Arena is a specialized branch of the LMSys Chatbot Arena platform that uses a blind-test battle mechanism — users submit programming tasks without knowing which models are involved, and the system presents outputs from two anonymous models side by side. Users vote for the better solution, and all results are aggregated into an Elo rating calculated using the Bradley-Terry model. It's one of the most recognized model capability leaderboards in the industry. Because of its anonymous mechanism, major vendors often use this platform to "soft-launch" new models before official release. Historically, GPT-4o, Claude 3.5 Sonnet, and other flagship models have all made brief anonymous appearances here.

One notable technical detail: both models have a knowledge cutoff in late 2025 to early 2026, strongly suggesting they are very recent frontier checkpoints. The training knowledge cutoff refers to the time boundary of a model's pre-training corpus — the closer the cutoff is to the present, the more recently the full pipeline (data collection, cleaning, training, and safety alignment) was completed. It's worth noting that there's typically a 3–9 month "data to product" lag between the cutoff and official release, covering instruction fine-tuning (SFT), RLHF, and red-teaming — making such a recent cutoff a strong release signal in itself.

According to Qwen community member "Tiewen," citing information shared by someone who spoke with Alibaba personnel at a Paris conference:

• Qwen 3.8 may be released soon, reportedly targeting performance beyond GLM 5.2;
• Qwen 4.0 may follow shortly after.

While no official timeline has been confirmed, historically, an anonymous model entering public testing typically signals that an official release is imminent. The "test before launch" strategy is also becoming a standard approach for vendors to gauge market reaction.

DeepSeek V4: Expanding Rollout, Public Launch Approaching

The most information-rich of the three leaks concerns DeepSeek V4. The model's limited rollout is reportedly expanding, with more users gaining access. Outputs leaked on social platforms show significant improvements in code quality, visual design, and overall reliability.

Understanding Gray Release: Gray release (also known as canary release) is a standard deployment strategy in internet engineering — a new version isn't pushed to all users at once, but gradually rolled out to 1%, 5%, 50%, and so on, keeping the "blast radius" of potential failures manageable. For large language models, new checkpoints often carry unpredictable regression risks in areas like inference latency, hallucination rate, and safety boundaries. A continuously expanding rollout, from an engineering standpoint, signals that the team's confidence in the new version's stability is linearly growing — a more direct "countdown to launch" signal than any official statement.

Windows 11 Clone Demo Makes a Splash

The leaker demonstrated a "Windows 11 clone" webpage generated by DeepSeek V4, with impressive results:

• Highly faithful reproduction of a real OS's appearance and widgets;
• Reasonable icons drawn for each application (e.g., the Edge browser);
• A built-in 3D neon racing game, Notepad, and Paint app.

In addition, a user had DeepSeek V4 build a game in pure HTML blending Minecraft and No Man's Sky styles, complete with full exploration mechanics, terrain environments, and a unified visual style. These demos suggest the model has developed a sense of "intent" when integrating multi-system and visual concepts, rather than just stacking technical components.

Three Signals Supporting an Imminent Public Launch

The leaker highlighted three notable release signals:

1. Expanding rollout — more users are being added to the testing program;
2. Chain-of-thought style changes — some users have noticed the new version may use phrases like "I'm going to" instead of the older "Let me," which can serve as a clue for identifying a new checkpoint. These phrasing differences aren't accidental: different teams develop distinctive writing styles in their chain-of-thought (CoT) training data, acting as a kind of linguistic fingerprint and an important indicator of a model's training origins;
3. Pricing notification commitment — DeepSeek has stated it will provide 24 hours' notice before any pricing changes take effect; if a release is indeed imminent, an official email could arrive at any time.

The Distillation Controversy: Did DeepSeek V4 Learn from Closed-Source Models?

The most controversial aspect of these leaks is the speculation that DeepSeek V4 may have "distilled" closed-source proprietary models.

Background on Knowledge Distillation: Knowledge distillation was first proposed by Hinton et al. in 2015. The core idea is to train a smaller "student model" using the output probability distributions of a larger "teacher model." In the LLM era, distillation has expanded to a broader context: using high-quality responses generated by closed-source flagship models (like GPT-4 or Claude) as fine-tuning data can significantly improve a model's instruction-following and reasoning capabilities at far lower cost than training from scratch. DeepSeek's earlier R1 model described a distillation process from its own models to smaller parameter versions in its technical report. However, using outputs from someone else's closed-source model for distillation is legally contentious: the terms of service of OpenAI, Anthropic, and others explicitly prohibit using API outputs to train competing models — which is precisely the crux of this controversy.

Researchers discovered several anomalies during testing:

• When handling complex 3D game or knowledge-base QA prompts, DeepSeek V4's outputs were sometimes strikingly similar to those of closed-source flagship models;
• Its chain-of-thought (CoT) style was "very different" from what users expect of DeepSeek, and instead resembled the reasoning style of closed-source models. CoT reasoning, proposed by Wei et al. in 2022, guides models to output intermediate reasoning steps before arriving at a final answer — different teams' CoT training data develops distinctive stylistic habits that form recognizable "linguistic fingerprints";
• When bio/cybersecurity-sensitive requests that would trigger closed-source model safety classifiers were included, output quality suddenly dropped significantly. Safety classifiers are filtering layers deployed in front of LLMs to identify and block high-risk requests — once triggered, the model refuses to answer or outputs a degraded response. This sudden quality drop is difficult to explain with purely self-developed model behavior.

One possible explanation is that certain complex prompts were being routed to proprietary models, and once a safety classifier was triggered, the system fell back to the self-developed model. The research team also noted that DeepSeek's routing behavior, output style, and CoT patterns were continuously changing during the investigation.

How to Interpret the Distillation Controversy

It's important to emphasize that these conclusions remain speculative — it cannot be definitively proven that DeepSeek is directly calling or distilling from closed-source models. However, the highly consistent behavioral patterns, classifier responses, and sudden quality drops make this theory difficult to simply dismiss. In fact, Anthropic and others have acknowledged that many labs engage in similar data collection practices.

The leaker's own take is telling: regardless of whether any lab is distilling from proprietary models, as long as they can deliver near-top-tier closed-source quality at dramatically lower cost, it's still a massive win for users. Under an aggressive pricing strategy, DeepSeek V4 has the potential to become one of the most disruptive coding models in the near term.

GLM 5.3/5.5: Zhipu Quietly Returns

The third piece of news is relatively brief: the Zhipu (GLM) team appears to have returned with a new model and has begun public testing of the next-generation GLM on internal platforms.

The new version may be GLM 5.3 or 5.5, with a small chance of jumping directly to version 6.0 (which the leaker is skeptical of). The only confirmed fact is that internal testing has begun. According to sources within the Zhipu team, a major model release is coming soon, potentially within the next quarter.

Summary: Chinese AI Models Enter a Period of Rapid Iteration

Taken together, these three leaks suggest Chinese AI models are entering a window of intense iteration:

• Qwen is testing the waters with two anonymous checkpoints, aiming to surpass GLM 5.2;
• DeepSeek V4, with its stunning demos like the Windows clone and cross-game fusion, combined with aggressive pricing, seems most likely to complete its public rollout first;
• GLM is following closely behind, maintaining its competitive pace.

Notably, while the distillation controversy remains unproven, it reflects a broader gray area around data sourcing that permeates the LLM industry — from IP questions around training data to the rise of "model watermarking" and "provenance tracking" technologies, the entire industry faces growing compliance pressure. For everyday users and developers, what may matter most isn't a model's "pedigree" but whether higher-quality outputs can be obtained at lower cost. If the three models launch as rumored, the competitive landscape of Chinese AI will undoubtedly see a pivotal reshuffling.

Key Takeaways