Apex Intelligence Releases Apex-35B Built by Its Own AI Research System
Apex Intelligence's first open model fine-tunes Qwen3.6-35B-A3B with help from an automated research system. It beats its base model on most research and agent benchmarks while activating only about 3B parameters per token.
Takeaways
- New lab Apex Intelligence released Apex-35B-V1.0, an Apache 2.0 fine-tune of Qwen3.6-35B-A3B.
- MoE model: 35B total parameters, about 3B active per token, 262K context, reasoning length chosen per turn.
- Apex Research, an automated system, generated training data, tuned the recipe and optimized GPU kernels.
- Big gains over base Qwen: SciCode 62.6 vs 37.5, AutomationBench 49.3 vs 10.0, BFCL V4 84.0.
- Still trails large models on BrowseComp and Terminal-Bench; text-only; key science benchmarks are internal.
- Runs on vLLM, SGLang and llama.cpp; GGUF builds range from 9 GB to 72 GB.
Apex Intelligence, a new lab founded by Tsinghua University assistant professor Yongchao Chen, has released Apex-35B-V1.0. The open-weight model fine-tunes Qwen3.6-35B-A3B for coding, tool use, deep search and computational research. Apex says its automated research system helped construct training data, refine the training recipe and optimize the infrastructure.
The release gives developers a concrete way to test whether automated research loops can improve a model sized for local deployment. Its Apache 2.0 license permits commercial use, and GGUF builds range from 9 GB to 37.8 GB. The evidence currently comes from Apex: the lab has not published a technical report, training data or enough detail to reproduce the training pipeline.
A 35B model for closed-loop research
| Component | Details |
|---|---|
| Base model | Qwen3.6-35B-A3B |
| Architecture | Sparse mixture of experts, unchanged from the base model |
| Parameters | 35 billion total, approximately 3 billion active per token |
| Context window | 262,144 tokens across input and output |
| Modality | Text only |
| License | Apache 2.0 |
The mixture-of-experts architecture routes each token through a small set of specialist subnetworks, so inference activates about 3 billion of the model’s 35 billion parameters. The long context can hold large repositories, research papers and extended tool histories, although input and generated output share the same 262,144-token budget.
Apex also describes the model as using “dynamic thinking.” It decides whether to produce an internal reasoning trace and how much computation to spend on each turn, without requiring a separate reasoning mode.
The intended workflow combines the model with an agent harness that supplies a shell, files, browsers, simulators or other tools. Apex-35B-V1.0 can then plan an experiment, write and run code, inspect the results and revise its approach. The model cannot execute those actions without the surrounding tools.
Apex Research shaped the fine-tune
Apex attributes much of the development work to Apex Research, its automated research system. The lab calls the process recursive self-improvement: the system identifies a weakness, proposes a change, runs an experiment, evaluates the result and carries successful changes into later rounds.