Limitations & known issues

Jev is strong inside its design space — and that space has hard walls. This page collects official limits, officially acknowledged issues, and the verification status of community disputes. Read it before committing.

Hard limits (official, verified 2026-09-22)

Limit	Fact	Consequence
No text generation	Typed answers only	No summaries, translation, or rewriting — hand those to an LLM
Text-only input	state accepts text (incl. serialized JSON)	Images and audio must be described in words first
≤ 64k tokens per request	state + longest question ≤ 32k	Long documents need chunking or retrieve-then-judge
Rate limits 250k tok/s, 1200 req/min	429-triggered, officially subject to change	Handle backoff at high concurrency
Output tokens free	Input $0.042/Mtok	Simple cost model, but long inputs still bill

Officially acknowledged issues

The vendor maintains a dedicated “Jev 1.13 jaggedness” page listing known rough edges of jev-1.13, stating most will be fixed in later versions (via the official llms.txt index, 2026-09-21). Publishing a defect list is commendable — but quote specific entries against the live page rather than copies (including this site’s).

What “can’t hallucinate” actually means

The marketing phrase refers to a type-safety property: answers can only take the values the question type allows — no fabricated out-of-format text the way an LLM can produce. It does not mean judgments are always correct. Picking the wrong option or mis-scoring is still possible, which is exactly why the docs push confidence gating. Reading it as “zero errors” was the most common misconception of launch week.

Community claims and their status

Claim	Nature	This site’s verification
“Vercel commands ran 5–18× faster after switching to Jev”	Media (TechCrunch) relaying Vercel	Not independently verified; vendor-reported
Homepage 193.6×/444.6× cost comparisons	Official	Blog self-annotates caveats: self-built workflows, West Coast evals, GPT-6 Astra & Fable 5.1 baselines
Theo et al. criticizing Jev-based context compaction	Community opinion	See Tutorials; the project (fast-jev-compaction) is in Ecosystem
Flood of fake Jev demos	Community observation	Debunked one by one in a Builder.io video (title verified); this site lists no unverified demos
“Jev is a Qwen wrapper”	Community rumor	No evidence; unverified and not endorsed here. The vendor states its own RLCD training

When Jev is the wrong tool

Our checklist (source: this site’s editors):

The answer space is open (copywriting, naming, free-form QA) → use an LLM
You need multi-step reasoning chains (proofs, complex planning) → Jev decides in single steps; chaining gets expensive
The judgment is fully rule-based (regex, SQL) → you don’t need a model
Errors are costly and no human review is possible → be careful; at minimum keep confidence thresholds and audit logs

Conversely, if your workload is “high-frequency, narrow judgments, wants probabilities, wants low latency,” it may fit — see Use cases.