[ Learn ] # Jev limitations: what it can't do Jev's limits from TypeSafe's own docs: no text generation, no audio or video, a fixed context window, English first. What to use instead for each. Published 16 October 2026By Jev Hub [Back to Learn](https://jevhub.io/learn) [ TLDR ] - Jev is a decision model, so it cannot write text, explain its answers or work through long chains of reasoning. - It does not accept audio or video, and image input is still experimental. - By default it reads 64,000 tokens per request, and English gives the best accuracy. - None of this is a bug: each limit follows from what Jev is built to do. ## The four big limits Four things Jev can't do. Each limit follows from Jev returning only typed decisions. The right column is what TypeSafe's documentation points to instead. ## Jev's limits at a glance | | The limit | What to use instead | | --- | --- | --- | | Writing text | Jev does not generate text. TypeSafe's introduction says: "No text generation, no parsing." Every answer is a choice, a score or a yes/no probability. | A language model for the writing. Let Jev decide whether writing is needed at all. | | Explaining an answer | An answer comes with probabilities and, for choice and score questions, a confidence number. It does not come with an explanation. | Use the confidence number to send unsure answers to a person or a stronger model. | | Multi-step reasoning | Jev is not designed for extended reasoning. TypeSafe suggests splitting such questions into narrower, separate ones. | Small questions combined in your code, or a reasoning model for the planning. | | Audio and video | Not accepted. The models page says: "It does not accept audio or video." | Convert the audio or video to text or structured fields first. | | Images | Accepted only behind an opt-in header. TypeSafe labels images "Dreamer-ready", meaning not ready for production, and says the limits may change. | Trials only, until TypeSafe calls it production-ready. | | Input length | By default 64,000 tokens per request and 32,000 per question. An opt-in header raises both to 262,000, which TypeSafe says is slower and less accurate. | Filter in code and send only what the question needs. | | Languages | English is the primary training language and gives the best accuracy. Other languages, including Chinese, Japanese and Korean, are "handled but not equally well". | Test with your own language before relying on it. TypeSafe says Jev 1.14 improves non-English most. | | Customising the model | Jev is not fine-tuned with customer data, and every account uses the same weights. TypeSafe says it does not train on customer requests or responses. | Put your specifics in the question's instructions and criteria, and test edge cases. | From TypeSafe's documentation pages for the introduction, models and images, read on 16 October 2026. Limits can change between versions, so check the pages before you build. ## What can't Jev do? Jev, the decision model from TypeSafe AI, cannot write text, explain why it chose an answer, work through long multi-step reasoning, or accept audio and video. Image input exists but is experimental, and the model reads a fixed amount of text at a time. These are the Jev limitations that matter before you build on it. They are not defects. Jev returns a typed answer with a probability, and that narrow job is the reason it is fast and cheap. The one-line version: if the output has to be a sentence, Jev is the wrong tool. ## Does Jev generate text? No. Jev does not generate text. TypeSafe's introduction to Jev puts it plainly: "No text generation, no parsing." Each answer is a choice from options you define, a score on a scale you describe, or a yes/no probability. That is why Jev does not replace ChatGPT or Claude for writing replies, summaries or code. The common pattern is to let Jev decide, then call a language model only for the items that need words: Jev reads every email and says which ones need a reply, and a language model drafts those replies. ## Can Jev explain why it chose an answer? No. Jev returns the answer and how likely each option is. For choice and score questions it also returns a confidence number, which shows how concentrated the probabilities are. It does not return a written reason. TypeSafe's advice is to use that confidence number as a gate: act automatically when it is high, and hand the item to a person or a stronger model when it is low. ## Can Jev reason through several steps? Not well, and it is not meant to. TypeSafe says Jev is not designed for extended reasoning, and recommends breaking a hard question into narrower questions and combining the answers in your own code. If the job is genuine planning or a long chain of logic, use a reasoning model for that part. Jev can still make the small decisions around it, such as which tool to call or whether to escalate. - [Where Jev struggles](https://jevhub.io/where-jev-struggles) ## Can Jev read images, audio or video? Audio and video: no. TypeSafe's models page says Jev does not accept either, so convert them to text or structured fields first. Images: yes, but only as an experiment. Both Jev 1.13 and Jev 1.14 accept images behind the header `TypeSafe-Danger-Zone-Images: true`. TypeSafe labels the feature "Dreamer-ready", which means not ready for production, and says the format and limits may change. Images must be PNG or JPEG, up to 16 MiB each, with each side between 280 and 1,568 pixels. TypeSafe states there is no separate charge for images: they count as input tokens. - [Jev 1.14: what changed](https://jevhub.io/updates/jev-1-14-images-box-and-longer-context) ## How much text can Jev read? (Jev context window) By default, Jev reads up to 64,000 tokens per request, and up to 32,000 tokens for any single question together with the input. With the opt-in header `TypeSafe-Danger-Zone-Long-Context: true`, both limits rise to 262,000 tokens, but TypeSafe says long context is slower and less accurate than the defaults. In practice, shorter input gives better answers. The fix TypeSafe recommends is to filter in your own code and send only what the question needs. ## Does Jev work in languages other than English? Yes, with a caveat. English is Jev's primary training language and gives the best accuracy. TypeSafe says other languages, including Chinese, Japanese and Korean, are handled but not equally well. It describes Jev 1.14 as small improvements across the board with the biggest gain in languages other than English; that is the company's own claim and it has not published numbers for it. If your input is not in English, test a sample of real examples before you depend on the answers. ## Can you fine-tune Jev? No. TypeSafe says Jev is not fine-tuned or adapted with customer data, and that all accounts share the same model weights. It also says it does not train on customer requests or responses. To make Jev fit your case, write precise instructions and criteria for each question and test the edge cases. The question design guide covers how. ## How are these limits different from where Jev struggles? The limits on this page are boundaries of the design: things Jev is not built to do. Where Jev struggles is a different list, TypeSafe's own account of weak spots inside the work Jev does do, such as counting, date comparisons, long noisy input and text written to steer the model. - [Where Jev struggles, and what to do instead](https://jevhub.io/where-jev-struggles) ## Our read: when not to use Jev Our read is that these limits are the product, not a flaw in it. A model that only decides can be fast, cheap and predictable; one that also writes, explains and plans cannot. The mistake is not picking Jev. The mistake is asking it for a sentence. Skip Jev when the output has to be written text, when you need the reasoning laid out, when the input is audio or video, or when an exact calculation is needed (that belongs in code). Use it for the high-volume yes/no, pick-one and rate-it decisions that sit in front of those steps. ## Common questions No. TypeSafe's introduction says Jev has no text generation. Every answer is a choice, a score or a yes/no probability. Use a language model when you need written text. No. It returns the answer, the probability of each option and, for choice and score questions, a confidence number. It does not return a written explanation. By default 64,000 tokens per request and 32,000 per question. An opt-in header raises both to 262,000 tokens, which TypeSafe says is slower and less accurate. Audio and video are not accepted. Images are accepted behind an opt-in header, but TypeSafe labels the feature as not ready for production. Yes, but English gives the best accuracy. TypeSafe says other languages are handled but not equally well, and describes Jev 1.14 as improving non-English the most. No. TypeSafe says Jev is not fine-tuned with customer data, all accounts share the same weights, and it does not train on customer requests or responses. Not for writing, reasoning or coding. Jev only returns typed decisions. Many teams use Jev for the high-volume decisions and a language model for the writing. ## Sources Written by Jev Hub from the sources below and checked against them on 16 October 2026. Where a figure is a company's own claim, the article says so. - [TypeSafe docs: introduction to Jev](https://docs.typesafe.ai/introduction) - [TypeSafe docs: models and limits](https://docs.typesafe.ai/models) - [TypeSafe docs: images (read 16 October 2026)](https://docs.typesafe.ai/images) - [TypeSafe docs: Jev 1.13 jaggedness](https://docs.typesafe.ai/model-jaggedness/jev-1.13) - [TypeSafe blog: Jev: For The Builders AND Dreamers (14 October 2026)](https://typesafe.ai/blog/jev-1-14) ## Related content - [Jev benchmarks](https://jevhub.io/jev-benchmarks) - [What is Jev?](https://jevhub.io/what-is-jev) - [How to use Jev](https://jevhub.io/how-to-use-jev)