Reflex Instinct 0.6B
A 0.6B decision model for browser agents: from a task and a page's accessibility tree it picks the next operation and the element to act on in one forward pass, with a probability per option and no generated text. LoRA on Qwen3-0.6B plus a pointer head; English web pages.
A bare Qwen3 model with no LM head; a pointer head scores options that are positions in the input, and any number of questions batch into one pass. Specialised for WebArena and BrowserGym accessibility trees rather than general decisions. Needs the repo's serve.py, which downloads Qwen3-0.6B and applies the adapter and head, and serves Jev's /v1/systemone wire format. Calibration comes from temperature scaling; training labels come from an LLM explorer and are noisy. On the maker's held-out NNetNav steps, element top-1 is 41.6% and the full step 31.0%. The maker's own run of JevBench's public items (a third-party benchmark) gives easy 91.7%, standard 66.7% and hard 28.8%, below that tier's chance rate. The companion Reflex Reason 2B writes reasoning before answering, so it is not listed. Unrelated to other models named reflex.
What it decides
- choice — picks one option from a set
- score — places the input on an ordered scale
- noul — answers a yes/no question with one probability
At a glance
| Parameters | 0.6B |
| Base model | Qwen/Qwen3-0.6B |
| Maker | Atlas AI (Atlas-AI-research) |
| Released | 2026-10-02 |
| License | apache-2.0 |
| Reported accuracy | 62.1% |
| Reported latency | 162 ms p50 / 452 ms p90 in-process on an RTX 5070 |
Get the weights
pip install systemonemodels
systemone pull atlas-ai-research/reflex-instinct
The files are served from the maker's Hugging Face repository, Atlas-AI-research/reflex-instinct-0.6b, and verified against the checksums recorded here.