TypeSafe AI released Jev 1.14.0 in preview on Oct. 14, under the model name jev-preview1. The company calls the update small. Most tasks should get a tiny bit better, and multilingual work should gain the most, TypeSafe wrote1. The bigger news sits behind three opt-in flags TypeSafe calls the danger zone: images, a new box type and longer context1. TypeSafe's post also published its first internal benchmark charts. Jev 1.14.0 leads them, on a test TypeSafe built, with rivals from Microsoft and OpenAI left off14.
Disclosure: AI Lately uses Jev to tag its pieces and to screen Both sides pairs before a person reviews them.
Two Audiences, One Model
TypeSafe frames the release as a split between two kinds of customer. "To really care about the prod builders, we should be shipping models that meet our quality bar for intelligence and have been battle-tested to the limits of what we can do internally," the post says1. The other group it calls dreamers. Those builders want experiments, and TypeSafe says it lacks a full picture of what will delight them1.
Its answer is a set of switches in the API headers. Each feature turns on only when a developer sends a header such as TypeSafe-Danger-Zone-Images: true1. TypeSafe stands behind the features and warns against shipping them. "But as a warning, we think there is still way too much jaggedness for us to vouch for these for production use cases!" the post says1.
What the Danger Zone Holds
Images come first. Jev 1.13.0 could already read images, TypeSafe revealed1. Developers can now place them in the state, the instructions or the criteria of a request1. TypeSafe adds its own caution: every model, Jev included, does worse on images than on plain text1.
The box type returns a single bounding box, the rectangle around one object in an image. It is TypeSafe's first type built for work beyond simple decisions, which TypeSafe calls smart-if statements1. Box runs only on Jev 1.14.01.
Longer context is the third switch. Jev supports a longer input than TypeSafe allows in production, and the limit was a deliberate choice, the post says1. "We aren't your parents though, so while we strongly recommend limiting context length, you no longer have to," TypeSafe wrote1.
A fourth change sits in the settings. Organizations can let TypeSafe read their inputs and outputs to study model jaggedness1. "We will still not train on your queries, but this does make it easier for us to find issues and help debug your use cases," the post says1.
The Upgrade Schedule
TypeSafe set dates for the switch. The jev-latest name moves to Jev 1.14.0 a week after the post1. Jev 1.13.0 stays available for a while1. TypeSafe intends to give 60 days of warning before it retires Jev 1.13.0, and 30 days for Jev 1.14.01. A written deprecation policy has yet to appear; TypeSafe says one will come first1.
Every Number in TypeSafe's Charts
TypeSafe measured eight models on its own scale of intelligence, where higher is better1. Jev 1.14.0 scored 60.0, and Jev 1.13.0 scored 58.9, a gain of 1.1 points1. The best rival, kev-27b, scored 46.7, which leaves Jev 1.14.0 ahead by 13.3 points1. Below it came pplx-decider-v1.1-27b at 46.0, clef at 44.3 and clef-flash at 38.81. Kev-4b scored 36.9, and d1-3b scored 33.81.
Jev 1.14.0 scores 60.0 on TypeSafe's own intelligence scale, 13.3 points above the best open-weight rival
TypeSafe's internal intelligence score, higher is better, eight decision models; TypeSafe says the test likely favors Jev
Each bar is one model and its length is TypeSafe's intelligence score; the two magenta bars are TypeSafe's Jev versions and the gray bars are the open-weight rivals it chose to test.
Source: TypeSafe AI Blog [1]. Chart by The AILately.com Desk.
The numbers behind this chart
| Item | Value |
|---|---|
| Jev 1.14.0 | 60.0 |
| Jev 1.13.0 | 58.9 |
| kev-27b | 46.7 |
| pplx-decider-v1.1-27b | 46.0 |
| clef | 44.3 |
| clef-flash | 38.8 |
| kev-4b | 36.9 |
| d1-3b | 33.8 |
Calibration tells a buyer whether a model's confidence matches its hit rate. A model that says 90% should be right about nine times in ten. TypeSafe charted calibration error, where lower is better1. Jev 1.14.0 posted 1.6, down from 1.8 for Jev 1.13.01. The rivals ran from 10.3 for kev-4b to 35.0 for pplx-decider-v1.1-27b1. Clef came in at 11.2, kev-27b at 17.8, clef-flash at 19.1 and d1-3b at 19.91.
Jev 1.14.0's calibration error is 1.6, against 10.3 to 35.0 for six open-weight rivals
TypeSafe's internal calibration error, lower is better, eight decision models, as TypeSafe charted them
Each column is one model and its height is calibration error, the gap between stated confidence and actual hit rate; the magenta columns are Jev and shorter is better.
Source: TypeSafe AI Blog [1]. Chart by The AILately.com Desk.
The numbers behind this chart
| Item | Value | Note |
|---|---|---|
| Jev 1.14.0 | 1.6 | |
| Jev 1.13.0 | 1.8 | |
| kev-4b | 10.3 | |
| clef | 11.2 | |
| kev-27b | 17.8 | |
| clef-flash | 19.1 | |
| d1-3b | 19.9 | |
| pplx-decider v1.1 | 35.0 | Full name: pplx-decider-v1.1-27b |
A third chart set intelligence against cost per million queries, on a scale that runs from $50 to $5,0001. That chart leaves exact costs unprinted, so positions here are read off its grid. Both Jev versions sit at one point between $100 and $200, so the upgrade leaves the price where it was1. D1-3b sits just above $50, the cheapest model charted1. Clef-flash sits between $100 and $200 as well, and pplx-decider-v1.1-27b between $200 and $5001. Clef and kev-4b sit near $1,000, and kev-27b falls between $2,000 and $5,0001. Kev-27b, the strongest rival on intelligence, is the most expensive model on the chart1.
What the Charts Leave Out
TypeSafe printed a banner over every chart: "Warning: AI benchmarks. Please don't trust them."1. The post explains the test set. "We're including only open weight models because we guard our validation set very closely and are spooked out about the data policies of our model providers," TypeSafe wrote1. It adds that the results are "likely biased towards our models"1.
Two rivals sold through company clouds are missing from the page. Microsoft released Microsoft-Decision-1 on Oct. 9, at Jev's price of $0.042 per million input tokens, CellCog reported43. Microsoft's model is a post-trained version of Alibaba's open-weight Qwen3.5-9B, served through Microsoft's own cloud4. On Microsoft's own 36-benchmark table, it averaged 83.5% accuracy against Jev's 82.3%4. Jev led on calibration in the same table, 93.7 against 92.24. OpenAI's Decisions API, another rival CellCog tracks, is also absent from TypeSafe's charts14.
Read together, the two sets of numbers point the same way. Jev's clearest edge, on Microsoft's table and on TypeSafe's, is calibration14. Accuracy is where Microsoft claims the lead, and TypeSafe's charts leave that matchup unmeasured14.
The Case For and the Case Against
TypeSafe's case is that the danger-zone features are early, real and ahead of the field. "These are features that will be here to stay for future models, are (as far as we can tell) better than what exists out there, and will continue to get better by leaps and bounds," the post says1. TypeSafe also tells readers to check its claims themselves. "Don't believe it entirely. Always favor your own experimentation and evals with your use cases," it wrote1.
Anastasios Angelopoulos, chief executive of the model evaluation platform Arena, questioned the category when Jev launched. "It's unclear to me what makes these models different from standard 'zero-shot classifiers', which are relatively well-known technology," Angelopoulos said5. A zero-shot classifier sorts text into labels it skipped in training. The technique predates Jev. Microsoft built its rival on someone else's open weights within four weeks of Jev's launch, a fact that gives the question weight24.
What Each Side Needs
TypeSafe's case needs its lead to show up where buyers can check it. That means an outside leaderboard running Jev 1.14.0 beside Microsoft-Decision-1. It also needs the danger-zone features to graduate into production on the schedule its post implies14. Angelopoulos's case needs rivals built on open weights to keep matching Jev on accuracy at Jev's price45. Calibration is the number both sides can test today, because Jev leads on it in both companies' tables14.
By the numbers
- 60.0: Jev 1.14.0's score on TypeSafe's intelligence scale, against 58.9 for Jev 1.13.01.
- 13.3 points separate Jev 1.14.0 from kev-27b, the best rival at 46.71.
- 1.6: Jev 1.14.0's calibration error, against 35.0 for pplx-decider-v1.1-27b1.
- 60 days of warning come before Jev 1.13.0 retires, and 30 days for Jev 1.14.01.
- $0.042 per million input tokens: the price Microsoft-Decision-1 matched4.
- 83.5% against 82.3%: Microsoft's accuracy figures for its model and Jev4.
What to watch
- An outside run of Jev 1.14.0 against Microsoft-Decision-1 on JevBench or another public test4.
- TypeSafe's written deprecation policy, which it says will come before any retirement1.
- The day the jev-latest name points at Jev 1.14.0, a week after the Oct. 14 post1.
- Which danger-zone feature TypeSafe first clears for production use1.
Sources
- TypeSafe AI, "Jev: For the builders AND dreamers," TypeSafe AI Blog, Oct. 14, 2026, https://typesafe.ai/blog/jev-1-14
- Marina Temkin, "The maker of non-text AI model Jev valued at $7.5B just weeks after launch," TechCrunch, Oct. 9, 2026, https://techcrunch.com/2026/10/09/the-maker-of-non-text-ai-model-jev-valued-at-7-5b-just-weeks-after-launch
- Achint Srivastava, "Microsoft-Decision-1: Our model for fast decision-making," Microsoft, Oct. 9, 2026, https://commandline.microsoft.com/microsoft-decision-1-model-foundry/
- CellCog, "Microsoft-Decision-1: Price, Benchmarks, vs Jev," CellCog, Oct. 10, 2026, https://cellcog.ai/blog/microsoft-decision-1/
- Nicole Jeffrey, "Start-up valued at $200mn fields $10bn offers after Jev launch," On Time Brief, Sept. 25, 2026, https://ontimebrief.com/en/2026/09/25/start-up-valued-at-200mn-fields-10bn-offers-after-jev-launch/