Original title: Apexintelligence-AI/Apex-35B-V1.0
The first model from the new organization Apex Intelligence, based on Qwen3.6-35B-A3B (total parameters 35B, approximately 3B activated per token), continued training with proprietary data and research trajectories. The official recommendation is to use it within Claude Code (v2.1.280). It covers daily programming, Q&A, search, and agentic tasks, as well as scientific workflows such as planning computations in fluid dynamics and robot simulations, writing and running programs, and improving methods based on results; it autonomously decides whether to think and for how long in each turn. The official statement notes that their proprietary automated research system, Apex Research, was used during development to generate training data for weak areas, tune training recipes, and optimize GPU kernels and inference. Context window: 262,144 tokens (shared between input and output).
Why it matters Among open-source models of the same size, specialized scientific workflows involving "writing code to run experiments and then refining plans" are rare; with approximately three billion parameters activated per token, quantized versions are available for local deployment.
Rising score 26 · Likes 103 · Downloads 2566
Reasons
- Official table comparison against the base Qwen3.6-35B-A3B: SciCode main problems 62.6 vs 37.5, BrowseComp 69.4 vs 42.1, BFCL V4 84.0 vs 75.2, HLE 47.2 vs 36.2; LiveCodeBench v6 82.3 is slightly lower than the base model's 84.5.
- In the same table, much larger models like GLM5.3 and DeepSeek V4.1 Flash still score higher on most metrics; its positioning is specialization in scientific research and agentic tasks within its size class.
- The GGUF version provides BF16 and 28 types of quantization, with the smallest IQ1_S being approximately 9.0 GB.
Usage
- hf download Apexintelligence-AI/Apex-35B-V1.0 --local-dir Apex-35B-V1.0 (BF16 is approximately 72 GB); for quantized versions, see Apexintelligence-AI/Apex-35B-V1.0-GGUF.
- Use SGLang ≥0.5.22 or vLLM ≥0.32.0 to launch an OpenAI-compatible interface; use qwen3 / qwen3_coder for inference and tool call parsers, with MTP speculative decoding enabled.
- You must use the chat_template.jinja provided in the repository; do not replace it with the Qwen3.6 template; llama.cpp requires the --jinja flag.
Risks
- The training data is proprietary, and the card does not specify its scale or sources; "recursive self-improvement" is merely an official description, and the three benchmarks RoboRSI, Real2Sim, and CFD-Bench are internal proprietary benchmarks that cannot be independently verified.
- All scores are officially self-reported; full-context BF16 requires a single GPU with over 141GB or multi-GPU tensor parallelism. There are startup issues with SGLang and Transformers under older drivers (570). This entry has not been run.