Apex-35B Benchmarks: How Was This AI-Built Model Made?
Apex-35B was built with an automated research system. Here is how it was made, where it beats Qwen 3.6, Flint and larger models, and where it falls short.
For Engineering and Executives

How was Apex-35B made, and how does it compare to bigger models?
Apex-35B is the first model from Apex Intelligence. It is built on Qwen 3.6 and uses a sparse mixture of experts design: 35 billion parameters in total, but only about 3 billion active for each token. Part of its training was done by an automated research system called Apex Research, which helped create training data, run experiments, tune the training recipe and optimize GPU kernels. On research tests it beats its base model and Flint, a similar-sized model. It also edges past some much larger models on a few tests.
TL;DR
- Apex-35B is a mixture of experts model with 35 billion parameters in total and about 3 billion active per token, built on Qwen 3.6.
- The team used an automated research system, Apex Research, to help produce training data, run experiments, tune the training recipe and optimize GPU kernels.
- On research benchmarks it beats its base model and Flint, a model of similar size.
- A larger model still leads on most tests. Apex-35B gets ahead of GLM 5.3 on two research tests and DeepSeek on one of the hardest biology questions.
- It decides on its own when to think and for how long, and is released under Apache 2.
- In one hands-on review it fixed a real billing bug, cleaned a messy spreadsheet and solved a genetics problem. Its front-end design output was only so-so.
What is a sparse mixture of experts model?
A normal ("dense") model uses all of its parameters to process every piece of text. A mixture of experts model splits much of its network into groups of specialists and routes each token to only a few of them. Apex-35B has 35 billion parameters in total, but only about 3 billion do work on any given token.
For you, this means two things. The model can hold a lot of learned knowledge, and each response is cheaper to compute than the total size suggests. You still need enough memory to hold all the parameters. In the hands-on test, the 8-bit quantized version (Q8, a compressed format that trades a little precision for a smaller footprint) served through llama.cpp used just over 38 GB of VRAM on an 80 GB H100. A lower-precision quant would need less. The full-precision BF16 version needs more.
How did an automated research system help build it?
Most models are shaped by people who write datasets, launch training runs, read the results and adjust. Apex Research automates parts of that loop. According to the team, it helped with four jobs:
- Creating training data. Producing examples the model learns from.
- Running experiments. Trying variations and checking what improved.
- Tuning the training recipe. Adjusting the settings that control how training proceeds.
- Optimizing GPU kernels. Rewriting the low-level code that runs on the graphics card so training and inference are more efficient.
So the model was partly built by AI during research, which is why it is described as "self-evolving." That label needs care. The sources describe AI assisting a research process run by a team. They do not describe a model that rewrites itself after release. Think of it as automation inside the lab, not a model that keeps improving on your machine.
How does Apex-35B score on benchmarks?
The reviewer showed a benchmark table but the sources do not give individual scores, so this article does not quote numbers. The pattern was clear, though:
| Comparison | Result |
|---|---|
| Apex-35B vs. Qwen 3.6 (its base) | Apex clearly wins on the research tests |
| Apex-35B vs. Flint (similar size) | Apex wins |
| Apex-35B vs. larger models, overall | The bigger model still leads on most tests |
| Apex-35B vs. GLM 5.3 | Apex ahead on two research tests |
| Apex-35B vs. DeepSeek | Apex ahead on one of the hardest biology questions |
Treat this as a sign, not a verdict. Beating a base model on research tests suggests the extra training worked. Edging past larger models on a few tests shows a small-active-parameter model can compete in narrow areas. It does not mean Apex-35B replaces those larger models. Before you rely on it for anything important, check the published table and test it on your own tasks.
How did it do on real tasks?
The reviewer, Fahad Mirza, ran it locally with an agent tool called Hermes and tried four things.
A bug hunt in a full-stack app. The app was a solar installer's operations console. One screen told a customer in Pune the system would pay for itself in 1.8 years. Another showed a jagged, staircase-shaped electricity bill chart. The model was not told what the bug was. In about five and a half minutes it found the shared billing error, rewrote the function, rebuilt only the API service and added a regression test. Pune's payback became 2.7 years, and the 25-year savings fell from about $44,000 to about $30,000. The chart became a smooth line.
A single-file animation. The request was a rain-on-a-foggy-window page in one HTML file with no libraries. It took about 24 minutes and the model said it verified features like frame rate, merging and wiping. The features worked, but the look missed the brief: thick, dark trails instead of thin streaks, plus noisy square patches.
A messy spreadsheet. A year of online store orders had duplicates, four date formats, inconsistent country names and bad values. The model produced a cleaned file of around 1,400 orders, a log of every problem, a chart and a findings report. One finding: electronics bring in about half the revenue, but roughly one in five electronics orders is returned. Some chart labels overlapped.
A genetics probability problem. More on that next.
What did the genetics reasoning test show?
The prompt asked for an exact probability from a family history, a medical test that can miss, and two healthy children. No code and no tools were allowed, so the model had to reason alone. It thought for about five minutes, longer than the base model, and got all three answers right, including 66.7% for the woman in question. It worked step by step, noting that the brother's status made both of her parents carriers, and it avoided several traps built into the prompt.
It was not flawless. One figure, 1.98%, skipped a small correction. Even so, this test lines up with the benchmark claim that Apex does well on hard biology questions. The tradeoff is time: longer thinking means slower answers.
What does this mean for you?
If you are not technical, the takeaway is about direction. Labs are using AI to speed up the work of building AI, and that makes small, capable models that run on your own hardware more common. Open licensing (Apache 2) means companies can use these models without sending data to an outside service.
If your team handles sensitive data, a model that fits on a single high-end GPU and runs locally is worth a look. Spreadsheet cleanup, bug fixing and document analysis were all done in the hands-on test without a cloud service.
If you are evaluating models, do three things:
- Pick two or three real tasks from your own work, not public puzzles.
- Compare Apex-35B against your current model on them, and time each run.
- Check outputs for visible quality as well as correctness. The animation test shows a model can pass its own checks and still look wrong.
What can you build when a model is only part of the answer?
A model that cleans a messy order spreadsheet is useful once. If you need it every week, with a shared dashboard, logins and a clear record of what changed, you need a small app around it. With Remy, you describe that tool in plain English, such as an order-cleanup page where your team uploads exports, flags returns and sees a revenue chart. Remy builds the full-stack app, tests it and deploys it to a live URL. You own the code, and you stop passing a spreadsheet around by email.
Frequently Asked Questions
Is Apex-35B open source?
It is released under the Apache 2 license, which permits broad commercial use. Check the license text and the model's release page for exact terms and what is included.
What hardware do I need to run Apex-35B?
In the hands-on test, the Q8 quantized version took just over 38 GB of VRAM on an H100 with 80 GB, served with llama.cpp. The reviewer said it can run on smaller GPUs, and a more heavily compressed quant would need less memory. The full-precision BF16 version needs more.
Does Apex-35B beat larger models like DeepSeek and GLM 5.3?
Only on a few tests. It got ahead of GLM 5.3 on two research tests and DeepSeek on one very hard biology question. A larger model still led on most tests overall.
Does a "self-evolving" model keep improving after release?
Nothing in the sources says so. The self-evolving label refers to Apex Research, the automated system that helped build the model during development. The released model is a fixed set of weights.
Useful for
How we source, write, and correct our articles. Editorial standards
Related articles

Is Tencent Hy4 Flash Good Enough for Daily Work? Tested

Maxwell V1 Flash Review: Coding, Languages and Research Tests

LongCat 2.5 vs Claude Opus 5: Benchmarks and Real Tests
