$ cat wiki/entities/apodex.md
Apodex
Latest
- 2026-08-24
Apodex 1.1
- 2026-08-11
Apodex Discovery
- 2026-07-21
Apodex Frontier Program
Overview
A lab building what it calls "Discoverative AI" — AI that makes new discoveries rather than generating outputs from existing material — with the stated goal of being a "Heavy Duty Solver" for problems that lack standard answers and where "humans do not know where to start". Its site tagline is "Self-Evolving Heavy-Duty Solver" (source).
Founded and personally led by Chen Tianqiao, who is reported to have provided the initial funding (source).
This page exists because the lab published two system papers in fourteen days that this wiki holds — Apodex Discovery: Reality Benchmarks and Environments for Evaluating and Building Discoverative Artificial Intelligence (arXiv:2608.11341) (2026-08-11) and Apodex 1.1: Scaling Agentic Intelligence for Complex Work (arXiv:2608.23283) (2026-08-24) — and because the second was the top entry of the 2026-08-26 HuggingFace Daily Papers snapshot at 172 upvotes, 2.7× the next paper (source). Both arXiv abstracts name no organisation; the attribution here comes from the company's own blog and HuggingFace org, searched for directly (source).
Key People
- Chen Tianqiao — founder, and reported to lead the company personally (source)
- Simon Shaolei Du — Chief Scientist for Reasoning Models and Training (source)
- Beibin Li — Chief Scientist for Self-Evolution and Programming (source)
No further staff, headcount or founding date is established by anything read.
Models & Products
- Apodex Deep Discover (shipped form of Apodex-1.0-H) — the heavy-duty solver. Reported to coordinate up to 150 sub-agents over 15,000 steps in heavy-duty mode, with specialised agents examining a problem, cross-checking findings and auditing evidence before answering (source)
- Apodex-1.0-mini — on HuggingFace under the
apodexorg; a third-party model index describes it as 262k context, open source. The licence name is not given in anything read (source) - Apodex 1.1 and Apodex 1.1 Mini (35B) — announced 2026-08-24 with arXiv 2608.23283. Whether the 1.1 weights are released is not established; the paper describes the Mini as "locally deployable", which is not the same statement (source)
No model page exists for any of these, because no benchmark figure for any
Apodex model appears in anything read — not a score, not a context window for
1.1, not a price. A model page written from that would be five unknown rows and
a name.
Recent Activity
- 2026-08-24: Apodex 1.1 published (arXiv 2608.23283) — "working capability" proposed as a unit of measurement, scaled along Environment Scaling and Agentic Coordination Scaling, with an AgentOS holding task state and provenance. Claims "the leading performance band" with a substantially smaller model than many frontier systems, and a 35B Mini. No number is given for either claim → Apodex 1.1: Scaling Agentic Intelligence for Complex Work (arXiv:2608.23283) (source)
- 2026-08-11: Apodex Discovery published (arXiv 2608.11341) — 561 industries surveyed, 423 problems assembled, 20 released; scored by HDS6 (Tools, Repair, Alternatives, Coherence, Evidence, Scope) independently of final-task success. Reported AAV capsid design 7% above published state of the art → Apodex Discovery: Reality Benchmarks and Environments for Evaluating and Building Discoverative Artificial Intelligence (arXiv:2608.11341) (source)
- 2026-07-21: Apodex Frontier Program announced — selected research institutions, academic labs and deep-tech startups get up to $100,000 per month in compute credits, plus access to the heavy-duty solver and dedicated engineering support (source)
Strategic Position
The two papers are one programme split across evaluation and system, and that is the position: Apodex Discovery builds the environments and the scoring, Apodex 1.1 builds the solver those environments measure. Owning both halves is what lets a small lab make frontier-band claims — and it is also why the claims cannot be checked from outside. Nothing read says whether Apodex 1.1 is evaluated on Apodex Discovery's own environments.
The differentiating bet is size. "Leading performance band with a substantially smaller model", plus a 35B Mini offered as locally deployable, puts the lab on the side of Post-Training Scaling's claim that capability has stopped arriving in parameter count. It is the same wager Z.ai stated as a law on 2026-08-20 and Apodex Discovery: Reality Benchmarks and Environments for Evaluating and Building Discoverative Artificial Intelligence (arXiv:2608.11341) stated as an architecture — with, so far, no published figure from this lab to test it against.
The commercial shape is unusual for a lab this new: a compute-credit programme aimed at scientific researchers and deep-tech startups rather than a developer API or a chat product. That is a distribution strategy for a solver whose unit of work is an investigation, not a request.