Apex Intelligence releases Apex-35B-V1.0, a Qwen3.6-35B-A3B derivative, with open weights on Hugging Face

Apex Intelligence, a Chinese AI startup founded in June 2026, released its first model, Apex-35B-V1.0, on October 28. The weights are available on Hugging Face. The model is built on Alibaba's Qwen3.6-35B-A3B and was further developed through mid-training and post-training aimed at coding, agent workflows and general tasks, as well as research work in AI, computational science and robotics simulation, as reported by Pandaily.

Pricing not yet disclosed. Context window and training cutoff date are not given in the Pandaily report.

Built with an automated research system

The company's pitch centers on how the model was made rather than its size. Apex Intelligence says its automated research system, Apex Research, took part in generating training data, running experiments, designing the training plan, and optimizing GPU kernels and the inference system.

In one example the company gives, Apex Research found that the model was weak at computational fluid dynamics tasks and analyzed where its reasoning fell short. It then generated and verified new training data that was fed into later training. According to the company's internal testing, that loop raised the model's CFD-Bench score by 32.1 points.

Apex Intelligence also says AI-driven optimization raised training throughput by 45% to 60% and made decoding 1.6 to 1.8 times faster.

Benchmark claims

All benchmark results published with the release come from the company's own evaluation. On that basis, Apex Intelligence claims Apex-35B-V1.0:

• beats both Qwen3.6-35B-A3B and Recursive SuperIntelligence's Flint on all six research benchmarks it reports;
• outperforms GLM-5.3 on FrontierScience and PaperBench;
• exceeds DeepSeek-V4.1-Flash-1017 on the human-difficult split of BioMysteryBench;
• posts the highest BFCL V4 score among the models it compared.

On two robotics tests, RoboRSI and Real2Sim, it reports 44.5 and 21.55. The much larger GLM-5.3-Flash scores 52.4 and 21.4 on the same tests.

Real2Sim asks a model to rebuild a working physics simulation from a single photograph: identify the objects, choose simulation assets with similar visual and physical properties, and recreate their layout so robot actions can be tested. RoboRSI measures how well a model controls simulated robots by writing code and revising it based on visual feedback.

Founder's framing

The company was founded by Chen Yongchao, an assistant professor at Tsinghua University's College of AI. He frames the work as recursive self-improvement, the idea of AI systems helping to build their successors. He told Tencent Technology that AI has made clear progress in carrying out research but still lacks research taste, the judgment of which problems are worth pursuing.

What this means

The release is a test of a specific claim: that an automated agent loop can find a model's weak spot, produce verified training data for it and measurably improve it. The CFD-Bench example, a 32.1-point gain, is the company's main evidence, and it comes from internal testing. The same applies to every comparison against Qwen3.6-35B-A3B, Flint, GLM-5.3 and DeepSeek-V4.1-Flash-1017. Open weights on Hugging Face mean outside researchers can check these results independently.

The robotics results are the most concrete data points: on Real2Sim, the company reports 21.55 against 21.4 for GLM-5.3-Flash, while on RoboRSI it trails at 44.5 against 52.4. Chen's own comment that AI still lacks research taste is a stated limit on how far the approach goes today.