HomeAI models
Together: Tev1 4B Experimental
Together
•
Tev1 4B Experimental
•
Released 
September 2026

Together
Tev1 4B Experimental

A 4B decision model that returns one option letter for routing, classification and policy checks, not open-ended chat.

Modality:
Text
model ID
togethercomputer/tev1-4b-experimental
Start for free
arrow black icon
View API Docs
arrow white icon

Output Speed *

N/A
tok/s

Intelligence Index *

N/A
/ 100

Context Window *

32768
tokens

Input price

0.252
Anytoken

Output price

0
Anytoken
A 4B Decision Model That Answers With a Single Option Letter Tev1-4B-experimental is a Together AI fine-tune of Qwen3.5-4B. You give it context, a question and 2–24 options, and it returns one answer letter. It targets routing, classification and policy checks, not generic chat. It is an experimental, narrow-purpose release, not a general-purpose LLM. It suits support-ticket triage, rule application and intent labeling, where application code maps the returned letter to an action. Test Tev1-4B-experimental on your own routing data via AnyAPI.ai

Performance

Single-Letter Decisions at 4B Scale: Narrow Scope, Unproven Accuracy

Tev1 handles one task: given state, a question and 2–24 labeled options, it returns exactly one letter for your software to map to a key. It was trained on 37,840 examples using LoRA fine-tuning, covering language classification, policy decisions and routing. Together's saved dev results are 880 of 1,000 on the main decision set, but they come from reused development benchmarks, not untouched tests. Treat it as a cheap, deterministic classifier stage that needs validation on your own labeled data.

Benchmarks

Tev1-4B-experimental on the Jev Decision Index: Roughly Half of Jev

On the Jev Decision Index 0.3.1, which weights 20% public benchmarks and 80% private tests, Tev1-4B-experimental scores 34.14. Jev scores 60.11. Calibration error is 0.104 against 0.074 for Jev, and lower is better. Together's own results are development numbers and are not independent. Use the index as the more conservative signal for general decision tasks, and measure accuracy on your own labeled tickets or policies before replacing a larger model.

Output Speed

*
N/A
tok/s

Intelligence Index

*
N/A
/ 100

MMLU *

Broad world knowledge and problem-solving
0
%

GPQA *

PhD-level scientific reasoning across physics, biology, chemistry.
0
%

HLE *

Adherence to multi-step structured instructions.
0
%

LiveCodeBench *

Tool-calling reliability in long agentic loops.
0
%

Technical Specifications

What the model supports

The model has 4.7B parameters and a 32.8K-token context window. It takes text input and returns a single option letter. Together recommends temperature 0, max_tokens 8 and thinking disabled, and the endpoint does not inject these settings automatically. The practical limit is input length: the training recipe used a 2,048-token sequence limit, so keep state short despite the larger advertised window.
Verified Specifications — 
Tev1 4B Experimental
*
Input modalities
Text
output modalities
Text
Context window
32768
 tokens
Maximum output tokens
8
Reasoning
No
Knowledge cutoff
September 2026
Pricing (standard)
0.252
 AnyTokens in
 / 
0
 AnyTokens out

Limitations & Trade-offs

Where Tev1 4B Experimental falls short

1
Closed-set output only. The model is meant for routing, classification and policy checks, not generic chat. In a regular chat it tends to reply in prose. It returns a letter, with no explanation and no documented score or probability primitive. Use a general instruction-tuned model for summaries, extraction, open-ended answers or anything that needs reasoning shown to the user.
2
Short effective input. The advertised context is 32,768 tokens, but the training recipe used a 2,048-token sequence limit, and the longest training example was about 1,500 tokens. Long documents, large option sets with long descriptions and multi-document RAG inputs fall outside the tested range. Chunk or summarize first, or use a long-context model.
3
Unproven accuracy. Together's saved results are reused development benchmarks, not untouched final tests. On the third-party Jev Decision Index it scores 34.14 against 60.11 for Jev. The model is also labeled experimental. For high-stakes decisions such as refunds, compliance or safety routing, add human review, a none option and your own labeled holdout set, or use a stronger model.
4
Manual configuration and licensing caveats. Temperature 0, max_tokens 8 and thinking disabled are not injected automatically, so a default request may not return a clean single letter. The license for the fine-tuned weights was still being finalized at last check. Review license terms before building a self-hosted product on the weights.

Best-Fit Workloads

Where this model earns its place

01

Support ticket routing

‍
Together's own example sends a billing complaint with billing, technical and sales options.The training mix included 6,000 routing decisions. The fit is a front-line triage step that assigns teams or queues in a single short call. Keep ticket text short, add a none option for ambiguous cases, and escalate low-confidence tickets to a larger model.

02

Policy and rule checks

‍
The largest training slice was 13,500 programmatic policy-rule examples. That makes it suited to questions like whether a return falls inside a stated window. The repo example disables thinking and constrains the answer to the supplied option letters. Its perfect policy-transfer score uses synthetic policies, so test it against your real policy wording before relying on it.

03

Intent, sentiment and taxonomy labeling

‍
Training data included Banking77 intents, SST-5 sentiment levels, AG News categories and research-paper taxonomy decisions. It suits high-volume labeling of short texts into a fixed label set, where the output is one letter and input tokens dominate the workload. It is not suited to open-vocabulary tagging or long-document classification.

04

Template for custom decision models

‍
Together describes fine-tuning your own classifier in about 25 minutes, behind an HTTP endpoint that answers in the format your software expects. Teams can evaluate Tev1 first on real decisions, then retrain it on their own labels using the published recipe. This suits organizations with proprietary categories or policies that a generic hosted model does not handle well.

Pricing in anytokens via AnyAPI
Input
0.252
₳
Output
0
₳
Cache write
—
₳
Cache read
—
₳
Get API access
arrow black icon
Read Docs

on this page

PerformanceBenchmarksTechnical SpecificationsLimitations & Trade-offsBest-Fit WorkloadsIntegrationFAQ

Integration

Access Tev1 4B Experimental via AnyAPI.ai

Access Tev1 4B Experimental through AnyAPI.ai using a unified API built for multi-model AI applications. Integrate Tev1 4B Experimental without maintaining a separate provider-specific connection, and keep the flexibility to test, switch, or combine models as your application requirements evolve.

01

One API integration

Access Tev1 4B Experimental and other AI models through the same API workflow instead of maintaining separate integrations for every provider.

02

Easy model switching

Test Tev1 4B Experimental against alternative models or switch models as your performance, capability, or cost requirements change without rebuilding your application around another provider API.

03

Flexible for production

Use Tev1 4B Experimental from experimentation through production while keeping your AI stack flexible as workloads, traffic, and model requirements evolve.

04

Multi-model applications

Use Tev1 4B Experimental for the workloads where it performs best and combine it with other models for tasks that require different capabilities, performance, or efficiency.

Start free
arrow black icon
Read Docs

Frequently Asked Questions

Answers to common questions about integrating and using this AI model via AnyAPI.ai

Tev1-4B-experimental is an experimental 4B decision model from Together AI, a supervised fine-tune of Qwen3.5-4B. It chooses one option from a structured state, question and list of choices. It is not a general-purpose chat model.

OpenRouter lists a 32,768-token context window, and Together lists 32.8K. The training recipe used a 2,048-token sequence limit, so short inputs are the safe range. Output is a single letter, and Together recommends max_tokens of 8.

No, it is not designed for chat. It is intended for routing, classification and policy checks. In a regular chat it tends to reply in prose, and it is trained to return one option letter with no explanation.

On a third-party index it is well behind Jev. The Jev Decision Index 0.3.1 gives Tev1-4B-experimental 34.14 and Jev 60.11. Together's own results are reused development benchmarks, not untouched final tests, so validate on your own data.

Together recommends temperature 0, max_tokens 8, thinking disabled, and a system prompt that tells it to treat state as data and return only the option letter. The endpoint does not inject these settings, so set them in every request. Then map the returned letter to your semantic key in application code.

* Benchmark data source: Artificial Analysis artificialanalysis.ai