SYLIN

Apex

Model · Language & code models

promising

Small open science-and-agents model that improves on its Qwen base; one reviewer so far

First model from Apex Intelligence: sparse MoE built on Qwen 3.6, 35B total parameters, about 3B active per token

AccessOpen weightsNeeds~71.9 GB; smallest build ~38 GBDownloadabout 72 GB full; Q8 needs about 38 GB of VRAMSize36B parametersLicenseApache 2.0

Why we say so

Fahd Mirza, the only hands-on tester, found it fixed a real billing bug, cleaned a messy spreadsheet and solved a hard genetics problem. His visual coding result was weaker, and he notes it thinks far longer than the base model. The benchmark wins over Qwen 3.6 and Flint are Apex Intelligence's own, and the maker admits larger models lead most tests.

Evidence

  1. Nov 18
    Supports 262K-token context, decides on its own when and how long to think, released under Apache 2.0
  2. Nov 18
    Built partly by an automated research system, Apex Research, that helped create training data, run experiments, tune the recipe and optimize GPU kernels
  3. Nov 18
    Via Hermes Agent, found and fixed a hidden billing bug in a full-stack app in about 5.5 minutes, adding a regression test
  4. Nov 18
    Clearly beats the Qwen 3.6 base model on the research tests in the benchmark table
  5. Nov 18
    Beats similar-size Flint on the benchmark table's research tests
Show all 7
  1. Nov 18
    Ahead of GLM 5.3 on two research tests and DeepSeek on one hard biology question; larger models lead most tests
  2. Nov 1
    Qwen-based 35B for scientific research and robot-control code; full 72 GB, smallest quant 9 GB; beats base Qwen 3.6

Cost to run

Alternatives

Made from