The author tests the Apodex 4B-SFT and 35B mini models, finding the 4B-SFT surpasses other 4B models in multi-hop search tasks without hallucination, and notes the design philosophy of separating answer checking from generation.
Weekend project writeup. The Apodex collection went up on HF a couple days ago and I grabbed the small ones to see what the fuss was about. For context these are the open releases from their 1.0 launch, the mini at 35B-A3B and the smol SFT line at 0.8B, 2B and 4B. The big 397B and the heavy mode are API only, so this is just the local stuff. I mostly lived in the 4B this weekend and only gave the 35b mini a quick spin since it is painful on one card. What makes them a bit different from a normal small model is they are trained to run as a search agent. Plan a query, call tools, then check their own work before answering, instead of one shot chatting. I wired the 4B-SFT into my own little ReAct harness with a search tool and threw a few multi hop questions at it, the kind where the answer is buried three links deep and most small models just confidently invent something. Rough impressions on my box, a 3090, running the 4B in fp16 via vLLM and the 35B mini through transformers with aggressive CPU offload since the full weights are still 35B on disk even though only about 3B are active per token. The offload gets it running but it is slow enough that I only use it for one off questions, not back to back. The 4B-SFT is genuinely better at not hallucinating the final hop than other 4B class things I have tried. The official claim is it beats every open 30B class model on BrowseComp and BrowseComp-ZH, and while I cannot reproduce a full benchmark at home, on my handful of questions it was clearly punching above its weight. For day to day local stuff the 4B in vLLM is what I actually reach for, the mini is overkill on one card. One annoyance, there is no official gguf that I can find, so I converted the 0.8B and 2B myself for llama.cpp and just kept the 4B in vLLM. If someone has a clean quant of the 35b mini, please drop it. The part I find interesting is less the scores and more the design idea, that the thing checking the answer should not be the same context that produced it. Apodex is one of a few groups pushing that lately and it is nice to see it show up in models small enough to run on one card. weights are in the apodex/apodex-1 collection if you want to play. will report back if the gguf conversion of the bigger one stops being cursed.
Apodex is an open-sourced deep research harness and AI model, which is a finetune of Qwen 3.5 35B A3B, achieving performance comparable to frontier models with only 3B active parameters.
Apodex 1.0 is a self-evolving AI system post-trained on Qwen3.5, achieving SOTA on BrowseComp, DeepSearchQA, and HLE-text. Its 4B mini model outperforms 30B-class models, with an AgentOS runtime for task orchestration. Open weights available.
Apodex 1.1 mini achieves competitive performance on agent benchmarks with significantly fewer parameters than larger models like Kimi K2.6, highlighting deployment cost advantages.
The article highlights the impressive performance of DeepSeek V4 Flash with Antirez Dwarfstar 4 on a high-RAM Mac, noting its superiority over other AI models and the reduced need for larger systems.