AnswersPublished20 sources
Apex-35B-V1.0: An AI-Trained Qwen Model, and What Its Benchmarks Show
Apex 35B V1.0 is an open weight model built on Qwen3.6 35B A3B, with weights on Hugging Face. The model is positioned for coding, tool use, general tasks, and scientific research, including computational science and robotics simulation.
Published byEdited with GPT-6 LunaImages generated with GPT Image 2
Research answer
Key takeaways
- Apex 35B V1.0 is an open weight model built on Qwen3.6 35B A3B, with weights on Hugging Face.
- The model is positioned for coding, tool use, general tasks, and scientific research, including computational science and robotics simulation.
- Apex reports leading Qwen and Recursive Flint on many of its tests, but the available materials do not provide comparable figures for training throughput or decoding speed.
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: What is Apex Intelligence’s Apex-35B-V1.0, released on October 28, 2026, and how do its Hugging Face weight availability, Qwen3.6-35B-A3B fo. Article summary: Apex-35B-V1.0 is Apex Intelligence’s first model, released on October 28, 2026: an open-weight, Qwen-based model intended for general-purpose use and research, rather than a foundation model trained from scratch.[2][4][6. Topic tags: general, general web, user generated. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks, charts with fa
Apex-35B-V1.0 is Apex Intelligence’s first model: an open-weight system built on Alibaba’s Qwen3.6-35B-A3B and further developed through mid-training and post-training. Its standout claim is not that it was trained from scratch, but that Apex Research helped generate targeted data and use evaluation results to guide improvements. The company reports a 32.1-point gain on its internal CFD-Bench; that result, like the model’s broader benchmark claims, has not been independently established in the sources cited here.2
17
21
What Apex-35B-V1.0 is—and where to find it
The model is designed for general-purpose use and research. Its model description lists coding, tool use, question answering, and deep search alongside research in AI, computational science, and robotics simulation. The release materials also describe support for a Claude Code harness.2
21
The model weights are listed on Hugging Face under Apexintelligence-AI/Apex-35B-V1.0. An official GGUF repository is also available for compatible local runtimes.21
23
How Apex Research reportedly improved CFD performance
Apex’s reported workflow uses model errors as a signal for further training. In the computational-fluid-dynamics (CFD) example, Apex Research analyzed evaluation errors, identified reasoning problems, generated and verified new scientific training examples, then added those examples to later training.17
28
Apex attributed a 32.1-point improvement on its internal CFD-Bench to this targeted-data loop. This is a points gain, not a percentage increase. The sources describe the result as an internal evaluation, so it should be treated as a company-reported finding rather than an independent benchmark result.17
28
Apex also describes its automated system as contributing to training-data generation and experiments; a company profile additionally attributes GPU-kernel optimization to the system.1
6 The materials cited here do not give comparable numerical results for training throughput or decoding speed, or enough measurement detail to assess those efficiency claims.
What the benchmark comparisons say
The comparisons below are reported by Apex or by coverage of its release. They are useful signals about the company’s evaluation results, not independently verified rankings.17
28
30
38
- Across 21 benchmarks: Release reporting says Apex-35B-V1.0 scored above Qwen3.6-35B-A3B on 20 benchmarks. Against Recursive SuperIntelligence’s roughly 40-billion-parameter Flint, it reportedly led on 18 and tied on one. The evaluations covered research, agentic tasks, coding, and reasoning.
17
- Research tests: Apex says the model outperformed Qwen3.6-35B-A3B and Recursive Flint across all six research benchmarks it evaluated.
28
30
- PaperBench and FrontierScience: A report gives Apex scores of 60.9 on PaperBench and 40.0 on FrontierScience, compared with 45.8 and 21.7 for Qwen3.6-35B-A3B, respectively. It also reports GLM-5.3 scores of 42.5 on PaperBench and 30.7 on FrontierScience.
- Other reported comparisons: Apex says the model outperformed GLM-5.3 on FrontierScience and PaperBench, exceeded DeepSeek-V4.1-Flash-1017 on BioMysteryBench’s Human-difficult split, and had the highest BFCL V4 score among the models it compared.
38 The BFCL claim is limited to that comparison set, not all available models.
- Robotics-related tests: Apex’s release reports scores of 44.5 on RoboRSI and 21.55 on Real2Sim. The model materials describe robotics simulation, robotic task execution, trajectory collection, and real-to-sim scene reconstruction as relevant capabilities.
21
18
What the evidence supports—and what remains unclear
The clearest example of Apex’s claimed automated-research approach is the CFD loop: identify errors, create and verify targeted examples, and retrain. That provides a concrete account of how automated analysis might contribute to model development. It does not, by itself, demonstrate autonomous or open-ended recursive self-improvement.17
28
The model is available as a Qwen-based open-weight release, and the company reports gains across internal evaluations. But the cited materials do not provide independently verified benchmark results or comparable measurements for training throughput and decoding speed. Those limits matter when interpreting performance claims or comparing the model with other systems.
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "Apex-35B-V1.0: An AI-Trained Qwen Model, and What Its Benchmarks Show"?
Apex 35B V1.0 is an open weight model built on Qwen3.6 35B A3B, with weights on Hugging Face.
What are the key points to validate first?
Apex 35B V1.0 is an open weight model built on Qwen3.6 35B A3B, with weights on Hugging Face. The model is positioned for coding, tool use, general tasks, and scientific research, including computational science and robotics simulation.
What should I do next in practice?
Apex reports leading Qwen and Recursive Flint on many of its tests, but the available materials do not provide comparable figures for training throughput or decoding speed.