Apodex-1.1-mini (Apodex AI)
One-line: Agent-specialized post-train of Qwen3.5-35B-A3B for long-horizon research/execution tasks (AgentOS multi-subagent orchestration, native function calling) — but its “benchmarks” are measured through a multi-agent Agent Team harness, and it publishes no standard coding evals (SWE-bench / Aider / BFCL), so single-model agent-loop suitability is unproven.
⚠ Benchmark provenance is lab-reported AND harness-level. The headline numbers (APEX-Agents 27.7, FrontierFinance 50.2, GDPVal 51.7 for
Apodex-1.1-mini w/ Agent Team) are measured on a multi-subagent Agent Team orchestration stack, not the raw single model; ReAct deployments score lower. No SWE-bench Verified, no Aider polyglot, no BFCL, no LiveCodeBench are published for the standalone checkpoint. These are not drop-in comparable to the SWE-bench/BFCL figures on other models — treat as directional only.
Release event: Apodex-1.1 family (mini 35B-A3B + full) publicized ~2026-08-17 (repo createdAt) / 08-24 blog + arXiv 2608.23283; model update 08-26. Apache-2.0. Built on Qwen3.5-35B-A3B, Qwen3.5 chat template (<think>..., <tool_call>), agentic RL for tool use, long-horizon task decomposition, and a built-in “Statement Review” verification step.
Coding benchmarks
| Benchmark | Score | Source |
|---|---|---|
| HumanEval | — | — |
| HumanEval+ | — | — |
| MBPP | — | — |
| LiveCodeBench | — | — |
| Aider polyglot | — | — |
| SWE-bench Verified | — | — |
| SWE-bench Multilingual | — | — |
Agentic / tool-use benchmarks
| Benchmark | Score | Source |
|---|---|---|
| BFCL | — | — |
| τ-Bench | — | — |
| ToolACE | — | — |
| GAIA | — | — |
| APEX-Agents | 27.7 (mini, Agent Team) | lab (arXiv, harness) |
| FrontierFinance | 50.2 (mini, Agent Team) | lab (arXiv, harness) |
| GDPVal | 51.7 (mini, Agent Team) | lab (arXiv, harness) |
Reasoning (context only)
| Benchmark | Score | Source |
|---|---|---|
| MMLU / MMLU-Pro | — | — |
| GPQA Diamond | — | — |
| MATH | — | — |
| AIME 2025 | — | — |
| Humanity’s Last Exam | 56.1 (Apodex-1.1 full) | lab |
Efficiency
- Active params: ~3B active (35B total MoE, base Qwen3.5-35B-A3B).
- Context length: 262,144 (256K).
- VRAM (fp16 rough): ~70 GB fp16 (35B total) → needs quantization (Q4 ≈ 20 GB) or multiple GPUs; ~3B active means the decode compute is cheap once the weights are loaded.
What makes it notable
Apodex-1.1-mini is explicitly engineered as an agent rather than a chat assistant — it decomposes tasks, coordinates parallel subagents, and produces verifiable deliverables rather than single-turn answers. The 256K context and ~3B active compute make it a plausible local long-horizon-agent runner. The catch is exactly what the skill’s “no-standard-benchmarks = exclude” rule flags: every published number is measured through a harness (Agent Team vs ReAct) with no standalone SWE-bench/BFCL result, so you can’t tell how the raw model behaves in your own agent loop. Worth tracking, but treat it as unverified for coding-agent use until a standalone eval exists.