Apodex 开了个新思路，模型不再单打独斗，而是训练出内部团队自己拆任务、检查答案、动态改稿。有开源版可以试试。
Apodex-1.0-H 是一种全新范式的深度研究模型，发布 open-weight 的 Apodex-1.0-mini 以及 0.8B、2B、4B 的 Smol 系列。模型原生像子代理团队工作：主代理分解查询，按需生成异步工作的专业子代理（研究、验证、事实核查、审计）。它通过 generate→verify→revise 循环动态改进答案，每一轮基于自身弱点评分并重写。验证过程使用独立子代理团队在多个类别上打分，避免自检盲区。您可在 HuggingFace 获取开放权重版本。
New paradigm for deep research models! Apodex-1.0…
New paradigm for deep research models!
Apodex-1.0-H is a new model that introduces a completely new way of working. Apodex comes in several flavors, including an open-weight Apodex-1.0-mini and 0.8B, 2B, and 4B Smol series of models.
This is what's new with Apodex:
1. The model works like a team 2. It dynamically improves its own work 3. It verifies every answer
First, the model works natively like a team of subagents to solve a task:
You have the main agent decomposing the query and spawning specialist sub-agents on demand. All of these subagents work asynchronously. You have agents for research, verification, fact-checking, and auditing, all working together.
I call this "natively" because the ability to decompose a task into multiple subagents was trained into the model.
Second, the model can dynamically improve its own answers.
As it works to produce an answer, the model grades it, notes where it's weak, and rewrites it based on that feedback.
loop: generate → verify → revise
Each pass of this loop learns from the last and improves on the answer.
The grader never sees a correct answer. It's judging whether its own work holds up, with nothing to compare it to.
Third, the model uses different agents to reason and verify.
The obvious hallucinations are easy to catch, but those where the answers look and feel right are much harder and more dangerous.
Models that check their own work have huge blind spots.
The verification process in Apodex uses a separate team of subagents to score answers on several categories to determine which answer actually solves the problem.
You can find the open-weight variants of Apodex on HuggingFace: https://t.co/JMLKlnq3av
To see Apodex-1.0-H in action, go here: https://t.co/YSnQtSq88T
Thanks to the Apodex team for partnering with me on this post.