Is Apex-35B-V1.0 state of the art?Pending benchmarks.Announced Oct 28, 2026; no independent results yet.
Apex Intelligence's first model: a general-purpose and research model fine-tuned from Qwen3.6-35B-A3B (35B total, about 3B active, mixture of experts) with further mid-training and post-training on proprietary data and research trajectories, partly produced by the lab's automated research system, Apex Research. Covers coding, tool use, question answering and deep search, plus computational research in AI, fluid dynamics and robotics simulation. Text in, text out; 262,144-token context; Apache-2.0 weights on Hugging Face.
The details so far
- Lab
- Apex Intelligence (first model)
- Announced
- Oct 28, 2026
- API
- Not yet
- Open weights
- Published on Hugging Face under Apache-2.0: BF16 safetensors (about 72 GB) and a GGUF repo (BF16, Q8_0, Q6_K, Q5_K_M, Q4_K_M)
- Parameters
- 35B total, about 3B active (35.95B in the safetensors)
- Base model
- Qwen/Qwen3.6-35B-A3B (fine-tune)
- Development
- Mid-training and post-training with input from Apex Research, the lab's automated research system (training data, experiments, training recipes, GPU kernel and inference optimization)
- Architecture
- Sparse mixture of experts; dynamic thinking (decides per turn whether and how long to reason)
- Context
- 262,144 tokens, shared by input and output
- Inputs
- Text in, text out
- Harness
- Built for use in Claude Code (v2.1.280 preferred); serves through SGLang, vLLM, TokenSpeed, Transformers and llama.cpp
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: ahead of its base Qwen3.6-35B-A3B on 20 of the 21 public benchmarks in its table (LiveCodeBench v6 is the exception, 82.3 against 84.5), and ahead of larger models (300B and 800B class) on selected benchmarks.
Lab claim: 73.4 on SWE-bench Verified (no internet), 65.2 on Terminal-Bench 2.1, 69.4 on BrowseComp, 84.0 on BFCL V4, 83.3 on GPQA Diamond, 87.2 on MMLU-Pro and 47.2 on Humanity's Last Exam.
Lab claim: 44.5 on its internal RoboRSI benchmark and 21.6 on Real2Sim, and 79 wins, 0 draws and 21 losses against Qwen3.6-35B-A3B on its internal CFD-Bench.
Sources
LABApex Intelligence blog: Introducing Apex-35B-V1.0apexin.com/apex-35b/ · Oct 28, 2026LABApex Intelligence announcement on Xx.com/apexin_ai/status/2115425524340478404 · Oct 28, 2026REPOApex-35B-V1.0 on Hugging Facehuggingface.co/Apexintelligence-AI/Apex-35B-V1.0 · Oct 28, 2026REPOApex-35B-V1.0-GGUF on Hugging Facehuggingface.co/Apexintelligence-AI/Apex-35B-V1.0-GGUF · Oct 28, 2026REPOapexin-ai/Apex-35B-V1.0 on GitHubgithub.com/apexin-ai/Apex-35B-V1.0 · Oct 28, 2026Compare when it lands
Unscored: its base model, Qwen3.6 35B A3B. A scored comparison appears once both are on a shared board.
Tracking continues until an independent board lists it.