Apodex 1.1 Launches with Open-Weights 35B Model, Agent Framework and Workbench
On August 25, ApodexAI released the Apodex 1.1 model family alongside an open-sourced toolchain — a combined launch of model, framework and workbench: the 35B-parameter open-weights Apodex 1.1 Mini, the FrontierAgent framework on GitHub, plus a web workbench and API platform. The launch is complete but the project remains early-stage; official figures claim roughly a 10-point eval gain when the same weights are run with the Agent Team harness, which, if replicated, would show harness design materially affects model performance.
Confirmed
- Apodex 1.1 Mini is a 35B open-weights model positioned as frontier-grade for professional work, finance and scientific research; per @KangwookLee, training code is open-sourced and weights/toolchain are available on HuggingFace.
- Per the official @apodex account, Apodex 1.1 extends executable environments to handle complex real-world tasks, training agents for long-horizon coordination with state maintenance and recovery for sustainable, verifiable progress.
- Per @xiaohu, it is a deep-research agent where a main agent orchestrates an async sub-agent pool (up to 150 per task), supporting long runs, automatic failure recovery, result verification and mid-run new requirements, executing actions like reading files and writing code.
- Per @DeryaTR, official examples show blind docking of Apigenin and Schisandrin B against the GLUT3 glucose transporter to demonstrate scientific-task capability.
- FrontierAgent is open-sourced on GitHub with a native CLI TUI, one-command install on macOS/Linux without pre-installed environments (Docker optional), supporting ReAct and Agent Team modes.
- Per @huangyun122, a Task Orchestrator dynamically decomposes tasks with sub-agents such as Evidence Review and Data Analysis running in parallel, streaming intermediate results into a Shared Task State; users can intervene or redirect at any time.
- A web workbench supports uploading documents, data, tables, images or code, alongside an API platform.
- Per official figures relayed by @SimonShaoleiDu, switching the same weights to the Agent Team harness yields roughly a 10-point eval improvement.
Unconfirmed
- The 10-point harness gain is an official claim; the benchmark name and raw scores were not given, and the molecular docking example is illustrative without full benchmark numbers — both await third-party verification.
Why it matters
- Bundling a 35B open-weights model with an agent framework and web workbench lowers the barrier to local multi-agent deployment, and open training code aids community study.
- The Shared Task State design with up to 150 sub-agents, failure recovery and mid-run requirement insertion targets the pain point of long-horizon tasks.
- If the same-weights, different-harness comparison holds, it shows harness design materially shapes model performance and merits community replication.
- Note that FrontierAgent has only 26 GitHub stars so far, indicating an early-stage project.
2026-08-25 ~ 2026-08-26 · 14 related posts
Primary sources
- Apodex 1.1 Released: Asynchronous Agent Teams & Open Benchmarks — KaiyuYang4 · 2026-08-25
- Apodex 1.1 Released: Open 35B Model and Agent Framework — SimonShaoleiDu · 2026-08-25
- [source] Apodex 1.1 mini: open 35B local model, harness swap adds up to 10 points — SimonShaoleiDu · 2026-08-25
- Apodex open-sources FrontierAgent: native TUI, one command, ReAct and Agent Team modes — SimonShaoleiDu · 2026-08-25
- Apodex orchestrates parallel sub-agents with dynamic task decomposition — huangyun_122 · 2026-08-25
- Apodex 1.1 launches with web workbench, open-source agent harness, and 35B model weights — huangyun_122 · 2026-08-25
- [source] Apodex 1.1 Scales Agentic Intelligence for Complex Real-World Tasks — apodex · 2026-08-25
- Apodex 1.1 released with impressive science benchmark scores — DeryaTR_ · 2026-08-25
- Apodex 1.1: Scaling Agentic Intelligence with Asynchronous Agent Teams — KaiyuYang4 · 2026-08-26
- Apodex 1.1 Open Sources Its Agent Framework: 150 Parallel Sub-Agents with Self-Healing — xiaohu · 2026-08-26
4 near-duplicate retellings: Kangwook_Lee · iamfakhrealam · _akhaliq · rohanpaul_ai