Jev vs SemIf
SemIf reads option probabilities out of an open LLM you already run — it does not train a new decision model. Jev is hosted and purpose-built. What that difference means in practice. Independent hosted gateway; not the official TypeSafe site.
What SemIf is
SemIf, by TheoLeeCJ, is MIT-licensed. An earlier name for the same project was OpenJev — check which repo you actually mean before integrating. It does not fine-tune anything: it takes a frozen open LLM, scores the options you declared from that model's logits, and returns a distribution. You can fit a temperature per workload to sharpen or flatten the result.
You bring the model
You supply the base model and a machine that can serve it — people run it from small GPUs up to larger Qwen-class models. Quality follows the model you attach: a strong base LLM gives decent distributions; a weak one, and no amount of logit reading fixes it. There is no separate Jev checkpoint inside SemIf.
Operations and cost
You operate the serving stack: capacity, upgrades, and calibration checks whenever you change base model or temperature. Marginal cost per decision is compute-only. The hosted Jev path charges 1 credit per 1,000 input tokens (from $10 for 100,000 credits), and the operations — model versioning, capacity, failure retries — are our payroll, not yours.
When it fits
Use SemIf when you already host an open LLM around the clock and want typed decisions without adding a second vendor, and the MIT license fits your project. Use Jev through Optio when the decision should be one hosted call with keys and prepaid credits — and when calibrated behavior matters more than owning the serving stack.
Strengths and honest limits
Jev's honest limits: you cannot fine-tune it or pin a version your compliance team signs off the way self-hosting allows, and text is the whole world to it. SemIf's hidden cost is subtler than hardware: every base-model swap re-opens the calibration question, and a temperature fit for last quarter's model is folklore, not data, this quarter.
Quick comparison
| Question | Jev | SemIf |
|---|---|---|
| Model | Purpose-built decision model, hosted | Your frozen LLM's logits, typed by SemIf |
| License | API access (proprietary) | MIT |
| Serving | Optio's endpoint, keys, credits | Your GPU, your server, your on-call |
| Calibration | Probability + confidence per question type | Logit-based; temperature you fit per workload |
FAQ
- Is SemIf a way to run Jev myself?
- No — SemIf does not ship Jev's weights and cannot reproduce its training. It types decisions from a different model's logits. The request shapes are similar; the answers are not the same model's answers.
- Both return probabilities — what differs?
- Provenance and support. SemIf probabilities derive from generative logits with a temperature you fit; Jev probabilities come from a model trained for bounded decisions, with confidence alongside. Practically: which distribution is stable enough to threshold is an empirical question — test both on ten of your own cases.
- Which costs less at small scale?
- At a few thousand decisions a day, hosted credits from $10 are usually cheaper than keeping a GPU warm around the clock plus the calibration work that base-model changes trigger.
- Is this the official Jev site?
- No. Optio is an independently operated gateway for calling Jev and is not affiliated with TypeSafe or with any alternative model mentioned on this page.