Is Talos-0.8B state of the art?Pending benchmarks.Announced Oct 22, 2026; no independent results yet.
Tractian's open decision model for industrial (OT) networks. It sits in the Talos gateway between AI agents, scripts or operator screens and industrial controllers, and scores each typed Modbus or OPC UA request in one forward pass: allow, serve from cache, defer or deny, plus nine other typed questions with calibrated probabilities. It does not generate text. A rank-16 LoRA adapter on Qwen3.5-0.8B-Base, trained on synthetic plant traffic. Weights and code under Apache-2.0.
The details so far
- Lab
- Tractian
- Announced
- Oct 22, 2026
- API
- Not yet
- Open weights
- Published
- Parameters
- 0.8B (Qwen3.5-0.8B-Base plus a rank-16 LoRA adapter)
- Base model
- Qwen/Qwen3.5-0.8B-Base (Apache-2.0)
- Method
- Rank-16 LoRA adapter on the base model; answers ten typed questions in one forward pass with calibrated probabilities (model card)
- Outputs
- action (allow / cache / defer / deny), needed_now, duplicate, servable, over_budget, contradiction, state_conflict, envelope, misbehaving, severity
- Protocols
- Modbus and OPC UA requests to industrial controllers
- Training data
- 161,602 synthetic labelled requests from simulated Modbus and OPC UA plants (model card)
- Out of scope
- Not a safety instrumented system and not an authentication mechanism (model card)
- Weights
- Hugging Face (tractian/talos-0.8b), PEFT adapter format
- Code
- github.com/tractian/talos (gateway proxy, docs)
- License
- Apache-2.0
- API
- None announced
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: small enough to run on a network switch, and polices AI-agent behaviour on OT networks.
Lab claim: on a synthetic held-out plant family, 98.2% of bad requests caught, 0.58% of good requests wrongly blocked and 93.6% four-way action accuracy (model card).
Sources
REPOTalos-0.8B model card on Hugging Facehuggingface.co/tractian/talos-0.8b · Oct 22, 2026REPOtractian/talos on GitHubgithub.com/tractian/talos · Oct 22, 2026LABAnnouncement on X (Canyon Dell'Omo)x.com/cany0n/status/2113315412314071448 · Oct 22, 2026Tracking continues until an independent board lists it.