What Tev1 is
Tev1 is Together AI’s fine-tune of Qwen3.5-4B into a Jev-style classifier: a System One model that reads a state and a closed set of options and returns one answer, rather than generated prose. Together published it as together/Tev1-4B-experimental on its serverless API and, alongside it, the full data recipe and a tutorial for training your own version.
Unlike Jev’s typed Choice, Score and Noul questions, Tev1 has one output shape: given a state, a question and 2 to 24 labeled options, it returns the letter of the option it picked. The repo’s example script, decide.py, sends a support ticket and four possible intents and gets back a JSON object with the chosen label and key. There is no separate Score or Noul response format documented.
How it was built
Together fine-tuned Qwen/Qwen3.5-4B with LoRA (rank 8, one epoch, learning rate 5e-5, a 2,048-token sequence limit) on 37,840 to 38,340 training examples, depending on whether you read the GitHub README or the tutorial article; the two give slightly different totals for the same dataset. The mix draws from MultiNLI, BoolQ, Banking77, AG News and SST-5, plus generated policy, routing and research-classification examples. Together says the run cost about $17 and took roughly 25 minutes on their fine-tuning service.
The README reports 880 of 1,000 correct on its main development set and 300 of 300 on a policy-transfer set. It is explicit that these are reused development benchmarks the team checked its own recipe against, not a held-out final test, and that endpoint access depends on your own Together account to reproduce.
What it’s good at
Short classification calls with a small, fixed option list: routing a support ticket to an intent, a yes/no/unsure read of a policy question, or picking a sentiment bucket, the same shape as the training datasets. Output tokens are free, so cost scales with the size of the state you send in.
The Jev Decision Index by multimodalart gives an independent number. Its edition 0.4.1, in the file generated on 12 November 2026, ranks entries by a Full score built only from private tests, 60% on the same skills and 40% in new domains, on a scale where 0 is random guessing and 100 is perfect. Tev1-4B-experimental scores 33.73 and Tev1-0.8B-experimental 12.47, against Jev 1.14’s 60.24. On the public benchmarks alone they score 29.51 and 13.10, against Jev 1.14’s 60.7. The headline changed in 0.3 and again in 0.4, so these numbers do not compare with the 0.2.1 figures this page gave before, or with the 0.3.1 Full scores of 34.14 and 12.02. Tev1-4B’s calibration error is 0.104, against Jev 1.14’s 0.080.
What it’s not for
Tev1 does not generate text, and it has no documented Score or Noul primitive, so anything needing a calibrated probability or a numeric level rather than a picked letter falls outside what’s published. The GitHub README calls its own benchmark numbers reused development results rather than an independent test, so treat any accuracy comparison with Jev as unverified.
Access today
The fine-tuning repo is MIT licensed and the weights are on Hugging Face under togethercomputer/Tev1-4B-experimental. The repo’s training and deployment scripts use Together’s fine-tuning and dedicated-endpoint services. The hosted together/Tev1-4B-experimental model is live on Together’s serverless API now, billed per the pricing above, with no waitlist mentioned in the launch materials. It is also listed on OpenRouter as togethercomputer/tev1-4b-experimental since 30 September 2026, served by Together at the same price.
Specifications
| System One Score | 46 from 1 of 3 boardsJev Decision Index 46. How the score works |
| Inputs | Text |
| Question types | Choice |
| Max Choice options | 24 |
| Score levels | Not documented |
| Questions per call | Not documented |
| Total context | 2,048 tokens |
| State budget | Not documented |
| Rate limit | Not documented |
| Licence | Not published as of 23 Oct 2026 |
| SDKs |
The repo's README says the model takes a state, a question, and 2 to 24 options and returns one answer letter. The saved training recipe used a 2,048-token sequence limit; the maximum number of questions per call and any separate state-token budget are not documented. There is no distinct Score or Noul output: every answer is a selected option letter. OpenRouter accepts a 32,768-token context for the model, well above the 2,048-token training sequence limit, and sets a maximum of 8 completion tokens. Inputs past what training covered fall outside the tested range.
Versions
- together/Tev1-4B-experimental, 23 Sep 2026, LoRA fine-tune of Qwen/Qwen3.5-4B, rank 8, one epoch, learning rate 5e-5. The repo's saved dev-set results: 880 of 1,000 on the main decision set and 300 of 300 on a policy-transfer set. The README says these are reused development benchmarks, not untouched final tests. Release notes
Use cases
What people use Tev1 for, one page per pattern.
Support inbox triage with System One models
Send a support ticket to Jev once with every question attached. Category comes back as a selected label, severity and frustration as numbers on scales you wrote, refund intent as a probability. Your code reads those values and decides what happens to the ticket.
Examples built with Tev1
The most-starred and most-viewed entries in the directory. All Tev1 examples.
Tev1 (Together AI): Jev-like classifier on Qwen3.5 4B
Data recipe and scripts to fine-tune a Jev-style classifier on Qwen3.5 4B from 37,840 examples for about $17. Together also serves the result, Tev1-4B-experimental, at $0.042 per million input tokens, and a step-by-step tutorial is on X.
23 starsTraining cost $17Serving price $0.042/M in

Jev Decision Index: Jev against 50+ open decision models
A leaderboard that runs Jev and 54 open System One models and clones through the same 120,000-question suite on one RTX PRO 6000, scoring accuracy and calibration. Jev leads the 0.2 edition at 51.67 on a chance-corrected scale, with AutoJev-27B close behind at 50.94.
Benchmarks
Published scores, each linked to its source. A row marked vendor-run was measured by the vendor itself, so read it as one side's result. The compare hub sets these beside the other models.
| Benchmark | Metric | Score | Run by | Date |
|---|---|---|---|---|
| Jev Decision Index 0.4.1 | Full score | 33.73 score | Independent | 12 Nov 2026 |
| Jev Decision Index 0.4.1 | Public score | 29.51 score | Independent | 12 Nov 2026 |
| Jev Decision Index 0.4.1 | Expected calibration error | 0.104 ECE | Independent | 12 Nov 2026 |
| Jev Decision Index 0.4.1 | Full score | 12.47 score | Independent | 12 Nov 2026 |
| Jev Decision Index 0.4.1 | Public score | 13.1 score | Independent | 12 Nov 2026 |
| Jev Decision Index 0.4.1 | Expected calibration error | 0.1257 ECE | Independent | 12 Nov 2026 |
| Jev Decision Index 0.3.1 | Full score | 34.14 score | Independent | 10 Oct 2026 |
| Jev Decision Index 0.3.1 | Public score | 29.51 score | Independent | 10 Oct 2026 |
| Jev Decision Index 0.3.1 | Expected calibration error | 0.104 ECE | Independent | 10 Oct 2026 |
| Jev Decision Index 0.3.1 | Full score | 12.02 score | Independent | 10 Oct 2026 |
| Jev Decision Index 0.3.1 | Public score | 13.1 score | Independent | 10 Oct 2026 |
| Jev Decision Index 0.3.1 | Expected calibration error | 0.1257 ECE | Independent | 10 Oct 2026 |
| Tev1 main decision set | Correct decisions | 88 % | vendor-run | 23 Sep 2026 |
- Jev Decision Index, Full score: Row: Tev1-4B-experimental. Counts 60% private same-skill tests and 40% private new-domain tests; public benchmarks no longer count. 0 is random guessing, 100 is perfect.
- Jev Decision Index, Public score: Row: Tev1-4B-experimental.
- Jev Decision Index, Expected calibration error: Row: Tev1-4B-experimental. Lower is better.
- Jev Decision Index, Full score: Row: Tev1-0.8B-experimental. Counts 60% private same-skill tests and 40% private new-domain tests; public benchmarks no longer count. 0 is random guessing, 100 is perfect.
- Jev Decision Index, Public score: Row: Tev1-0.8B-experimental.
- Jev Decision Index, Expected calibration error: Row: Tev1-0.8B-experimental. Lower is better.
- Jev Decision Index, Full score: Row: Tev1-4B-experimental. Counts 20% public benchmarks and 80% private tests; 0 is random guessing, 100 is perfect.
- Jev Decision Index, Public score: Row: Tev1-4B-experimental.
- Jev Decision Index, Expected calibration error: Row: Tev1-4B-experimental. Lower is better.
- Jev Decision Index, Full score: Row: Tev1-0.8B-experimental. Counts 20% public benchmarks and 80% private tests; 0 is random guessing, 100 is perfect.
- Jev Decision Index, Public score: Row: Tev1-0.8B-experimental.
- Jev Decision Index, Expected calibration error: Row: Tev1-0.8B-experimental. Lower is better.
- Tev1 main decision set, Correct decisions: 880 of 1,000 on the repo README. Together calls it a reused development benchmark, not an untouched final test.