Apex Intelligence Releases 35B Model After Four Months, Pushing AI-Driven Research Into Focus

China’s Apex Intelligence has released Apex-35B-V1.0, a 35B-parameter general-purpose model developed with automated AI research, according to coverage from 巴士的報 and Pandaily. The launch comes roughly four months after the company was founded, according to BigGo Finance, and puts a newly formed Chinese AI lab into a crowded model market with a distinct emphasis on research automation.

The release is notable less for a disclosed benchmark result than for the development method the company is presenting. Apex Intelligence describes the model as the product of AI systems that assist with research, while co-founder or executive Chen Yongchao has argued that AI self-improvement remains limited by a problem he calls “taste.” The comment points to a broader question for model builders: can automated research systems reliably decide which ideas are worth pursuing, or can they only optimize against objectives humans have already defined?

Because the available source material consists of article titles and summaries rather than full technical documentation, details such as the model’s license, training data, inference requirements, evaluation results, and availability could not be independently confirmed from the supplied evidence.

What Apex Intelligence released

The central product is Apex-35B-V1.0, described by 巴士的報 as a general-purpose model and by Pandaily as a Qwen3.6-based model built with AI research agents. The two descriptions establish the model’s name, approximate scale, and claimed lineage, but do not provide a complete technical specification.

A 35B model sits in a size range that can be relevant to teams balancing capability against serving cost. In practice, the usefulness of a model at this scale depends on factors that are not disclosed in the source material: quantization options, context length, hardware requirements, latency, tool-use support, and performance on coding, reasoning, multilingual, and domain-specific tasks.

That missing information matters for potential users. A model release can be important to researchers while remaining difficult for enterprises to adopt if deployment documentation, safety testing, and commercial terms are unclear. For now, the confirmed news is the release itself and the company’s positioning around automated research, not a demonstrated advantage over established models.

Automated research is the product story

The strongest differentiator in the coverage is not simply that Apex Intelligence has released another language model. It is that the company says the model was built with AI research agents or automated AI research. Pandaily’s description specifically links Apex-35B-V1.0 to Qwen3.6 and AI research agents, while the 巴士的報 headline emphasizes automated AI research.

The phrase can cover several different activities, including experiment planning, code generation, data preparation, evaluation, literature review, and iteration on training recipes. The supplied evidence does not establish which of these functions Apex Intelligence used, how much autonomy the systems had, or how human researchers supervised them. Those distinctions are important because an agent that automates routine experiments is materially different from one that proposes and validates new model-training methods.

For AI builders, the announcement reflects a shift in competition from model architecture alone toward the systems used to create models. Research automation could help smaller teams run more experiments with limited staff, but it also creates new failure modes. An agent may optimize a narrow benchmark, select misleading evidence, or repeat a plausible but unproductive research direction at high speed. Reliable evaluation and human review therefore remain central even when the research loop is partly automated.

Chen Yongchao’s “taste” problem

BigGo Finance frames the launch around a comment from Chen Yongchao that AI self-improvement is stuck on “taste.” In this context, taste appears to refer to judgment: recognizing which research question is valuable, which result is meaningful, and which apparent improvement is merely noise or benchmark overfitting.

That distinction is consequential. Automated systems are generally good at searching within an objective, but the objective itself may be incomplete. A research agent can generate more candidate experiments, yet still struggle to identify a useful problem, weigh conflicting evidence, or understand why a result matters outside the test environment. Those are not just engineering problems; they are selection and interpretation problems.

The comment also places a limit on the most expansive claims about AI self-improvement. If systems can automate parts of experimentation but cannot consistently choose high-value directions, the result may be faster iteration rather than independent scientific progress. Chen’s observation is an executive or researcher viewpoint reported by BigGo Finance, not an independently verified measurement of current AI capabilities.

Implications for model teams and buyers

For model developers, Apex Intelligence’s release highlights the potential value of building internal AI research agents around the training pipeline. Such systems could reduce the time spent on repetitive experiments and make smaller teams more competitive. But the value will depend on reproducibility, experiment tracking, evaluation quality, and the ability to detect when an automated process is optimizing the wrong target.

For enterprise buyers, the immediate question is not whether Apex-35B-V1.0 was produced with AI research agents. It is whether the resulting model is dependable for a specific workflow. Buyers would need evidence on privacy, deployment options, model behavior, tool integration, safety controls, and total serving cost before considering production use.

The launch may also increase pressure on larger labs and open-model developers to explain how much of their own research pipeline is automated. Yet the available coverage does not show that Apex Intelligence has surpassed competing models, achieved broad adoption, or introduced a new training method. Any such conclusion would go beyond the evidence currently available.

What to watch next

The next meaningful signals will be technical and operational. Apex Intelligence’s documentation should clarify whether Apex-35B-V1.0 is downloadable, under what license it can be used, and what hardware is required. Independent evaluations will be needed to test its coding, reasoning, multilingual, and instruction-following performance rather than relying on launch positioning.

Researchers should also look for evidence about the AI research agents themselves: which tasks they performed, what human oversight remained in the loop, how experiments were selected, and whether the process produced improvements that can be reproduced by outside teams. The company’s handling of safety evaluation and model updates will be equally important.

Finally, adoption signals should be treated cautiously. The supplied reports do not verify customers, usage volume, or integration into commercial products. Those details would provide a clearer indication of whether automated AI research is a practical advantage or primarily a development narrative.

Creati.ai perspective

Apex Intelligence’s launch is an early signal that model companies may compete not only through parameter counts and benchmark scores, but also through the productivity of the research systems behind them. That could matter substantially for new labs if automated experimentation lets small teams explore more options with fewer resources.

The “taste” problem is the more important takeaway. Faster experimentation does not automatically produce better research, and an AI agent’s ability to generate or test ideas is not the same as its ability to judge significance. Apex-35B-V1.0 will be easier to assess once the company publishes reproducible technical evidence, deployment details, and independent results.