Small research AI model claims to beat larger open models
The Apodex team says it has released a family of small deep-research models. The post says Apodex-1.0-4B-SFT beats open-source 30B-class models on BrowseComp and BrowseComp-ZH. The team says careful training data, not only model size, drives research ability. It says the weights are released under Apache 2.0.
Key points
- Apodex says it released 0.8B, 2B, 4B, and 35B-A3B models.
- It claims Apodex-1.0-4B-SFT scored above open-source 30B-class models on BrowseComp and BrowseComp-ZH.
- The team says careful data construction mattered more than parameter count for this result.
- It says the models kept close to their Qwen3.5 base models on general knowledge, math, instruction-following, and long-context tests.
- The weights are said to be available under Apache 2.0.
Quick term guide
- deep-research
- AI work where a model searches, reads, and combines information before answering.
- open-source
- Software whose code is shared publicly so others can inspect, use, or change it.
- 30B-class models
- Large AI models with roughly 30 billion learned internal values.
- BrowseComp
- A test that checks how well an AI can answer hard questions using web browsing.
- training data
- The collection of information used to teach an AI how to recognize patterns and answer questions.
- AI agents
- AI agents are AI tools that can carry out steps toward a goal, not just answer once.
- tool calls
- Times when an AI system uses another function, such as search or file access.
- tool call
- One time an AI agent uses a tool, such as search, calculation, or file reading.