Apodex 1.1: open-weight agentic models, and a 35B you run locally
Apache-2.0 weights: a locally-run 35B Mini plus an API-hosted full model, an open multi-agent harness, and leading scores on finance and research agent tasks.
Copy markdown
A model family built to finish work, not chat
Apodex 1.1 targets long-horizon agentic tasks — acting across file, search, and code environments with multi-agent coordination to deliver verifiable results rather than one-shot answers. It ships in two tiers, a full model and a 35B Mini, both Apache-2.0.
The 35B Mini runs on your own hardware
Apodex-1.1-Mini is a ~35B mixture-of-experts model (built on Qwen3.5-35B-A3B, ~3B active) with a 256K context, deployable via vLLM or SGLang behind an OpenAI-compatible endpoint. That's a self-hosted deep-research agent with no per-token bill and no data leaving your infra.
You get the agent harness, not just weights
Apodex open-sourced FrontierAgent, the multi-agent scaffolding the models run inside — the file, search, and code 'worlds' and the coordination loop. Clone it this weekend to stand up an autonomous research or finance agent instead of wiring the orchestration yourself.
It punches above its parameter count
The team reports Apodex 1.1 reaching the 'leading performance band' across professional, finance, science, and reasoning work despite a far smaller network than frontier systems — including 50.2 on FrontierFinance (top among tested systems) and 27.7 on APEX-Agents.
Two ways in: local weights or hosted API
Pull the Mini weights from Hugging Face for local runs, or hit the full model through platform.apodex.ai when you need the bigger network. The tech report details the architecture, and Apache-2.0 keeps commercial use open.