FULL STORY

Apodex 1.1 Launches with Open-Source 35B Model and Agent Framework

ApodexAI released the Apodex 1.1 model family on August 25, open-sourcing its toolchain as a combined model, framework, and workbench offering. Early hands-on reviews from multiple bloggers praised the research agent's task decomposition and multi-format data handling.

2026-08-25 ~ 2026-08-26 · 2 episodes · 16 posts

Episode 1 · Apodex 1.1 Launches with Open-Weights 35B Model and Agent Framework (2026-08-25, 12 posts)

On August 25, ApodexAI released the Apodex 1.1 model family along with an open-sourced supporting toolchain — a combined "model + framework + workbench" launch. Weights are available on HuggingFace, with an API platform and web-based workbench launching alongside.

Confirmed

  • Models: Apodex 1.1 Mini is a 35B-parameter, open-weights model designed for local execution, positioned as a frontier-level model for professional work, finance, and scientific research tasks; weights and toolchain are fully open source via HuggingFace.
  • Framework: FrontierAgent is open-sourced on GitHub, with a native command-line TUI — one command to run on macOS/Linux, no pre-installed environment needed, Docker optional; it supports ReAct mode and Agent Team mode (asynchronous multi-agent collaboration with task decomposition and parallel coordination).
  • Platform: The web workbench supports uploading papers, data, spreadsheets, images, or code and completing tasks via prompts, and an API platform is also provided.
  • Benchmarks: Per official figures relayed by @SimonShaoleiDu, switching the same weights to the Agent Team mode harness improved benchmark scores by about 10 points.

Unconfirmed

  • The "10-point gain from switching harnesses" is an official claim; the specific benchmark names and original scores weren't given in the posts, and the magnitude of improvement awaits third-party verification.

Why it matters

  • The release bundles a 35B-class open-weights model with an agent framework and web workbench, lowering the barrier to local multi-agent deployment.
  • If the "same weights, different harness" comparison holds, it suggests harness design significantly affects model performance — worth watching for community replication.
  • Note that FrontierAgent currently has only 26 stars on GitHub; the project is still early-stage.

Episode 2 · Research Agent Apodex Praised in Hands-On Tests (2026-08-25, 4 posts)

Multiple hands-on reviews praise Apodex, a research AI agent that decomposes complex tasks, writes code to process raw files, and delivers verifiable outputs with traceable sources. It handled changes mid-task without restarting and produced an 80,000-word course and a 60,000-word manual.