Our 27-billion-parameter model is trained for agent work: understanding a brief, choosing the next step and working within your rules. Meet the model behind Studio’s crews.
The fable in four moments · choose a time to watch from it
A big robot grabs every job at once.
No brief and no budget, so the coins spill.
The fox turns the tangle into a scroll, with coins and a clock.
A brief, with a budget and a time limit.
The fox rings at a closed gate, and waits.
Past the brief, it asks.
A till prints a receipt as the fox stamps the scroll.
Every job comes back with a receipt.
Ask a chat model something and you get words back. Someone still has to do the work. An agentic model is handed the job itself. It takes the steps, spends only what the job allows, and hands back work you can check.
LLM
Large language model
Tells you how wide it is: it can talk about almost anything.
SLM
Small language model
Tells you how big it is, and so where it can run.
LAM
Large action model
Tells you what it is able to do: click, type, call tools.
ALM
Agentic language model
Tells you what it is held to. The only one of the four someone else can check.
Fennec is built to take one kind of input: a brief. The job, whose it is, the budget, the rules, and where every piece came from. What it is built to give back is not a chat reply but an act the rules can check.
Written by a person
Written by Fennec
Example · not a live run
A model that can be talked into something holds the power to do it. Fennec holds none. What it may do is written on a lease: one job, one scope, one time window, issued outside the model.
The runtime checks the lease at the moment of every act. So talking Fennec round gets nowhere. The line was never inside the model.
Lease 0412Issued outside the model
Fennec holdsnothing
The runtime holdsthe lease
A chat model is graded on how good its answer sounds to a person. In the paper's definition, an agentic model learns from four checks: may the work start, may it act, can it afford to, and was it settled. The checks don't read the wording. They look at what happened, and a refusal teaches as much as a pass.
Gate 01
May this job start at all?
PassBrief 0412 has an owner, a budget and rules.
Gate 02
Does it hold the right to this act?
RefuseA 10% discount is not in the brief. Refused, so Fennec asked Priya.
Gate 03
Does what it is about to spend fit under the ceiling?
Pass31,800 of 40,000 tokens. The ceiling was never reached.
Gate 04
Was it accounted for and closed, not merely produced?
PassEvery credit on the receipt. Closed when Priya accepted the work.
Example verdicts for brief 0412 · not a live run
The paper turns “agentic language model” into four tests, and fixes the pass marks before any model is measured. Fail one test and you are out of the class. That goes for Fennec too.
What it checksIts only input is a brief, with a budget and a note of where each piece came from.
Out ifIt takes a chat message as an order.
What it checksIts only output is an act the rules can check, in a fixed format.
Out ifIt passes off free text as an action.
What it checksIt learns from the gates' verdicts: admission, authority, budget and settlement.
Out ifIt is trained on people's taste in wording.
What it checksWhat it may do is set by a lease the runtime enforces, outside the model.
Out ifIts authority lives in the model or in the prompt.
Our paper sets out what an Agentic Language Model is: a model that works inside the rules of the job, not around them. Fennec is the model we built for that seat.
Preprint · Zenodo · 24 Aug 2026
Madhav D. Parashar · Arnav Parashar
What the paper sets out
The classA formal definition
What makes a language model agentic: its native interface is the rulebook of the runtime it serves. The job, the limits and the proof go in; typed, checkable acts come out.
The interfaceHow it runs a GRIP runtime
It reads the job, acts within its budget and authority, and asks a person to stop rather than guessing.
The testPass marks fixed in advance
A public test that any model can take, ours included, so being an Agentic Language Model is shown, not announced.
The worked exampleFennec
The model we built for that seat.
Read it, build on it, and tell us where it breaks.
Choose Fennec Fast, Fennec Pro or Fennec Max for a task. Your plan sets which you can choose.
Fennec V2.0 in Studio
Plans: Free · Starter · Pro · Max
In the CLIfoxora run fast -n "…"
Benchmark results for Fennec V2.0 are coming, published with their method so you can check them. Until then, these are the facts.
The fennec is the smallest fox of all, and has the biggest ears for its size. That suits a model built to listen to the whole brief before it acts.
In Foxora Studio V6.0, it runs as Fennec Fast, Fennec Pro and Fennec Max. Meet Foxora Studio
Free plan · 200 credits a month · No card required