Apex Intelligence releases its first AI model, developed with automated research

Apex Intelligence released its first AI model on October 29, 2026, with downloadable weights and a claim about how it was built. The company says its automated Apex Research system helped design experiments, construct training data and optimize training. The resulting Apex-35B-V1.0 builds on an existing Qwen model rather than starting from scratch.

The 35-billion-parameter general-purpose model is intended for coding, question answering, deep search and workflows in which AI uses tools. Apex also positions it for computational research in AI, science and robotics simulation. Its release announcement makes two related claims: better results than similarly sized rivals on selected tests, and a development process that uses AI to find and address weaknesses.

A Qwen foundation, an automated development loop

Apex-35B-V1.0 starts with Qwen3.6-35B-A3B. Apex further developed that foundation through additional training stages, incorporating proprietary datasets and records of research work. The distinction is important: this release is a further-trained version of an existing model, while Apex Research is the system the company says helped improve it.

The company describes a feedback loop: identify a weakness, generate a hypothesis, run experiments and feed useful findings into the next training round. AI participated autonomously in data construction, experiment design, code execution, analysis of training results and optimization, according to Apex. Its description of self-improvement covers changes to data, training methods and computing infrastructure.

One example involves computational fluid dynamics, the study of how fluids move. After evaluations exposed failures, Apex Research analyzed reasoning gaps, generated and validated scientific training examples, and added them to later training rounds. Apex reports a 32.1-point improvement on its internal CFD-Bench. That figure measures progress on the company’s own test, not an independently established gain across scientific research.

The system also examined records of agents working through tasks. Apex says it identified overly long reasoning contexts and an imbalanced mix of training data as problems. It then explored different sequence lengths, proportions of short and long task records, and ways to group training work by token count.

Answer quality versus computing efficiency

Apex’s benchmark comparisons put the model against its Qwen foundation and a somewhat larger, 40B competitor. The company also says it is competitive with substantially larger models on selected benchmarks. These are company-run evaluations; they support a narrower claim about those tests, rather than a universal ranking of model quality.

A separate set of results concerns the machinery behind the model. Apex reports 45–60% higher training throughput after work on balancing GPU workloads, overlapping communication with computation, and grouping work dynamically. This measures how much training the infrastructure processes, rather than whether the model produces better answers.

For running the trained model, Apex reports 1.6–1.8 times faster decoding and a 30–35% reduction in time to first token under comparable workloads. Those measures describe response generation speed and the wait before output begins. The company attributes the gains to changes in attention processing, cached information and execution.

From the release to a local deployment

The model card lists 3 billion active parameters per token, a 262,144-token context length and approximately 72 GB of BF16 weights. The active-parameter figure describes the portion used for each token, not the total download size. The context length sets how much text the model can work with at once.

• Weights are available on Hugging Face for developers to download and run.
• Code, deployment guidance and usage examples are available on GitHub.
• Licensing terms, hardware requirements and supported inference frameworks are detailed on the project pages; the release announcement does not specify a license.