Tev1
Together AI's experimental Qwen3.5 fine-tunes (4B and 0.8B) that pick one lettered option from a state, question and option list.
Overview
Tev1 is an experimental decision model from Together AI, released as supervised fine-tunes of Qwen3.5 (the model card covers Tev1-4B-experimental; the list also mentions a 0.8B version and an Ollama build named tev1). It is trained to read a structured state, a question and 2 to 24 labelled options and return exactly one option letter, which application code maps back to a semantic key. The card stresses it is Jev-inspired but not a non-autoregressive Jev runtime: it keeps Qwen's normal language-model head.
Together published a walkthrough on training a decision model for $17 and the full data recipe and code. The card gives a recommended system instruction that treats state as data, and request settings with temperature 0 and thinking disabled. Development results show 88.0% on a 1,000-item decision set and valid single-letter outputs on all 1,300 test calls, but the card notes these are not an independent benchmark and that calibration, prompt injection and multilingual behaviour are not comprehensively tested.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Is the return within the window? | yes · no |
How it works
- Send the system instruction and a JSON decision with state, question and lettered options.
- Generate with temperature 0, thinking disabled, up to 8 tokens.
- Read the single option letter.
- Map the letter back to the option key in application code.
Strengths
- Small open model usable via Together API or locally.
- Published training recipe and low reported training cost.
- Consistently valid single-letter outputs in development tests.
Caveats
- Experimental; licence for the fine-tuned weights is still being finalized.
- Returns a letter, not calibrated probabilities.
- Development results only, without an untuned baseline.
Getting started
Call model together/Tev1-4B-experimental via the Together chat completions API with the recommended system instruction and temperature 0.At a glance
- Type
- Model
- Status
- research
- Language
- Python
- Runtime
- Together API / transformers / Ollama
- Has code
- yes
Use cases and integrations
How curated lists describe it
Together AI's experimental 4B and 0.8B supervised fine-tunes of Qwen3.5, released with a walkthrough on [training your own decision model for $17](https://www.together.ai/blog/how-to-train-your-own-jev). In Ollama as `tev1`.
AnotiaWang/awesome-decision-models
Frequently asked questions
Is Tev1 a Jev clone?
It is Jev-inspired, but it is a standard autoregressive fine-tune that outputs an option letter.
How accurate is it?
The card reports 88.0% on a development decision set and notes this is not an independent benchmark.
Can I train my own?
Yes, Together published a walkthrough and the data recipe and code used for Tev1.
Related projects
Mangaba AI's local, calibrated typed-decision model for Portuguese text, trained to optimize the Brier score; proprietary since v5.
ProjectproductionPython- blink★ 22
Open one-pass typed-decision model with a dependency-free C99 runtime that also builds to a 66 KB WebAssembly module.
ServicebetaC - Foq★ 11
Local open alternative to Jev using a 1.58-bit ternary 8B model to answer booleans, choices, scores and Pydantic extraction in ~25 ms.
ProjectbetaPython Runs the open Laya decision model entirely in a browser tab with WebGPU and ONNX Runtime Web: no server, no API key.
ServicebetaTypeScriptllama.cpp-based C++ server that turns any LLM into a System One decision model, reading answers from next-token logits.
ProjectbetaC++Three open 4B decision models for English, Turkish and German that serve the Jev /v1/systemone API on a CPU in about 3 GB of RAM.
ServicebetaPython