Published
Updated
Together prices Tev1 input at $0.04, and Ollama serves Choice, Score, and Noul
Opened October 7, 2026, Together's Tev1 page lists input at $0.04 per million tokens and free output. The September 23 post listed $0.042. The chat call returns one letter. Ollama serves tev1 on /v1/systemone.
Together AI posted tev1-4B-experimental on September 23, 2026. The post puts it on Together serverless at $0.042 per million input tokens, with output at $0, and says training cost $17.
Opened October 7, 2026, the model page lists input at $0.04 per million tokens and free output. The endpoint is together/Tev1-4B-experimental. The page says 4.7 billion parameters, a 32.8K context, and a release date of September 22, 2026. The sample is POST /v1/chat/completions. The instruction tells the model to return one letter from 2 to 24 options.
The Hugging Face card says this is a fine-tune of Qwen3.5-4B that still uses the language-model head. It should return one option letter. The development numbers on that card are 880 of 1,000 on the main set and 300 of 300 on policy transfer. The card says those sets were used while the model was built. It says the license for the fine-tuned weights is still being finalized. The base model is Apache-2.0. The training code is linked from that card at togethercomputer/tev1.
Ollama’s tev1 page, opened the same day, serves tev1 and tev1:0.8b at /v1/systemone. The page shows choice, noul, and score. Choice and score take 2 to 26 options. The page says the model was trained on 2 to 24. One request can hold 1 to 64 questions. The tag line prints a 256K window. A note on the same page says a prompt has to fit in about 2,000 tokens.
The same page prints a mean on 3,880 human-labeled decisions: Tev1 4B 73.3%, Tev1 0.8B 63.5%, Nimble 9B 75.7%, and Jev 1.13 at 76.0%. It says the Jev cell is Bespoke Labs’ published API run. That mean is already on the Nimble page. It is not Together’s 880 of 1,000 development set.
We did not call Together, and we did not run the Ollama tag.
This site's reading
Editorial notes evaluating claims against primary sources, contextualizing findings alongside related implementations, and defining technical terms.
Verify
Together AI, @togethercompute, September 23, 2026, 22:06 UTC, status 2102882216950763814. The post says tev1-4B-experimental is on Together serverless at $0.042 per million input tokens and $0 output, and that training cost $17. The reply one second later calls the release experimental and links the Hugging Face card. We did not open the quoted tutorial article.
On October 7, 2026 we opened https://www.together.ai/models/tev1-4b-experimental. The page lists input at $0.04 per million tokens, free output, endpoint together/Tev1-4B-experimental, 4.7B parameters, and a 32.8K context. The released line says September 22, 2026. The sample call is POST /v1/chat/completions. The instruction says to return one letter. We did not send that call.
The Hugging Face card, opened the same day, says the checkpoint keeps Qwen's next-token language-model head and should return one option letter. The development set is 880 of 1,000 and 300 of 300. The card says those sets informed training. It says the fine-tune license is being finalized. The base is Apache-2.0.
Ollama's tev1 library page, opened the same day, shows a curl to /v1/systemone with type noul, and tables for choice, noul, and score. It prints Tev1 4B at 73.3% and Tev1 0.8B at 63.5% on 3,880 decisions, beside Nimble at 75.7% and Jev at 76.0%. The tag line says 256K. A note on the same page says the prompt has to fit in about 2,000 tokens. We did not run Ollama.
Compare
Nimble's page already prints this Ollama mean. The 73.3% and 63.5% are that table, not Together's 880 of 1,000 development set. The September 23 price and the October 7 price are two readings of Together's own rate. Jev's listed input price on this desk stays $0.042. We did not call either host.
Terms
- Tev1
- Together AI's experimental Qwen3.5 decision model, in 4B and 0.8B sizes. The Together chat call returns one letter. Ollama's library page serves it as Choice, Score, or Noul.
- Development set
- Together's 1,000-item main set and 300-item policy-transfer set. The card says the mixture informed training, so the 88% and 100% are not an untouched benchmark.