What is Jev? Questions on TypeSafe's System One model

Short answers from stories already on this desk. Figures are claims from the people who published them unless a third party ran the same test. Start is a reading order.

What is Jev?

TypeSafe AI's first System One model, announced September 15, 2026. Software sends unstructured state plus typed questions. The model returns a Choice, a Score, or a Noul with probabilities. What Jev returns is the explainer. The launch story is Almeida's thread and TypeSafe's workflow table. The standing page is what Jev is. Jev Day 2026 is the October 14, 2026 release of jev-1.14.0, plus opt-in images, a Box type, and 262k context.

What is Jev Day 2026?

Jev Day 2026 is TypeSafe's October 14, 2026 release of Jev 1.14.0, plus opt-in images, a Box type, and 262k context. On October 16, at 08:19 UTC, one Box for every vehicle returned one rectangle, n = 3, and type array returned HTTP 400. Date headers 08:49:11 GMT to 08:49:40 GMT October 26, POSTs with jev-latest and jev-preview named jev-1.14.0. At 00:28 UTC October 16, TypeSafe's jev-latest still returned jev-1.13.0. Aviz Maeir scored Jev 1.14 Box at 82% on 180 scenes. Yamashita's Kyoto-ben 60 scored 0.833 on 1.14.0 against 0.817 on 1.13.0. The standing page is Jev Day 2026. The dated stories are jev-1.14.0, images, Box, and 262k context.

Can Jev read images?

Images need the header TypeSafe-Danger-Zone-Images: true. On October 16, 2026, at 08:19 UTC, one Box for every vehicle returned one rectangle, n = 3, and type array returned HTTP 400. TypeLLM's cookbook at 14.00 s prints six vehicle boxes, including Tesla's flying Roadster. A request without that header returned HTTP 400. Box is a Jev 1.14 type. The write-ups are images and Box.

Can a decision model like Jev hallucinate?

TypeSafe's September 15, 2026 launch post says Jev gives up string generation and can't hallucinate. The same post says schema matching is guaranteed, and that the 0% on its plot is not an empirical count. A Choice can still be the wrong option. Cloudflare's card prints RAGTruth hallucination F1 at 76.5 for Jev. Perplexity's 1,500-row RAGTruth cell is a different test, at 77.27% for Jev. The standing page is decision model hallucination.

What is OpenAI Jev, or ChatGPT Jev?

On October 23, 2026, Date headers 02:23:12 GMT and 02:23:13 GMT, two POSTs from this desk to the Decisions API returned HTTP 200 from gpt-6-luna. OpenAI Developers posted general availability on October 22, 19:07 UTC. The guide prices uncached input at $0.10 per million tokens and cached input at $0.01, with no output charge. Requests now hold up to 400 questions. OpenAI Jev and ChatGPT Jev are search names for that API. The write-up is OpenAI Jev. The September 29 posts are on the earlier page.

What is Clef?

On October 9, 2026 Cloudflare's blog prices Clef-flash at $0.038 per million input tokens and Clef-omni at $0.15. Clef stays at $0.24. The October 1 cards listed Clef-flash at $0.09 and a 65,536-token window. The October 9 blog lists a hosted 24k window for Clef-flash. The October 1 median row is 209.3 ms and 38.8 ms against a printed Jev cell of 524.1 ms. typesafe/jev on the same host is a different card. The write-ups are Clef and Clef-omni.

What other decision models work like Jev?

A catalog of hosted calls and open checkpoints is decision models like Jev. OpenAI's Decisions API is generally available on gpt-6-luna. Cloudflare's Clef weights are a separate card from typesafe/jev. Databricks ai_decide is a Beta SQL function. The docs do not name the model and do not print a price. NIRNAY 450M prints 0.8792 on 3,080 Banking77 rows, and the card says the Jev cell was not measured in that repo. WaterSheep serves POST /v1/systemone on port 8766 and prints 77.8% in distribution and 61.2% on held-out datasets, with no Jev column. StartLux-Decision v1.1 27B Instant prints Full score 69.12, first of 174 on Decision Index 0.3.2, README opened October 20. The October 3 self-run of edition 0.2.1 stays 63.88. Weights are CC BY-NC 4.0. Sieve-31B prints Full score 63.8 on that 0.3.2 board. Drex Max prints public 67.59 on Nace's 0.3.2 bar. basal-1.5 prints 72.12% on Werdykt v1, against Jev 1.13.0 at 81.64%. Decision 2.0 Vega prints 56.5 on its own Decision Index 0.2.1 run. The card does not print a Jev row. Vela 2.0's 9B card prints 41.63 on Decision Index 0.2.1 with Noul calibration on, and 41.09 with it off. The evaluation file prints Lux at 46.23 and does not print a Jev row. fsdecide scores 91.3% of 80 review items against jev-latest at 56.3%, and 60.5% against 90.5% on the authors' general set. JEV-27B-VL finishes 15 of 20 pick-and-place scenes. Its six-benchmark mean is 84.07 against a hosted Jev cell of 83.85, and that JevBench column is not the official board. Tev1, opened October 7, 2026, is $0.04 per million input tokens on Together's page. The chat call returns one letter. Ollama serves choice, score, and noul. Mercury Decide cites 14 decisions a second and prints no score. Opened October 8, 2026, the API board ranks it third at 72.4, and the free model page returned, priced Free. SubQ Snap 1.1 prints 65 on Artificial Analysis Decision Index. Docs name subq-snap-1.1 at $0.04 per million input tokens. Date headers 02:23:17 GMT to 02:23:20 GMT October 23, three POSTs from this desk returned HTTP 200. mammoth-1, posted October 21, prints Decision Index 0.4 Full 71.24 on the card. Opened October 23, live 0.4.1 Full 68.4 rank 3. OR-decision-4B, posted October 30, prints JevBench 193 of 231, 83.55%, with launch-chart Choice 79.7 against Jev 79.6. MiniCPM5-2B-Jev, posted October 28, prints 78.79% on JevBench, first among open models at or under 2B. BranchOne, posted October 29, returns Choice probabilities from Qwen3.5-0.8B-Base, median 0.305 s on ten short text questions. Microsoft-Decision-1, posted October 9, 2026, lists $0.042 per million input tokens and says it was 4.5 times quicker than Quyet-1.0-Large. The post does not print an accuracy. The Foundry catalog, opened October 10, says the weights are not distributed. The write-ups are ai_decide, NIRNAY, WaterSheep, StartLux-Decision, mammoth-1, OR-decision-4B, MiniCPM5-2B-Jev, BranchOne, SubQ Snap, Drex Max, Sieve-31B, Clairic, Datadog, basal-1.5, Decision 2.0, Vela 2.0, fsdecide, and JEV-27B-VL, Tev1, Mercury Decide, and Microsoft-Decision-1. TypeLLM scores Qwen3.8-27B at 228 of 231 public JevBench tasks with thinking, and 195 of 231 without it. The homepage prices input at $0.05 per million tokens. The hosted name typellm-latest is not that file.

What is Strands Decider?

Strands Decider 2B is an Apache 2.0 checkpoint posted October 1, 2026, with 1.9 billion parameters. The reference recipe, v19, scores 167 of 231 on JevBench's public tasks, at a median 115 ms per question on an RTX 3090. The repository dates that rank to the September 25 v1.4.2 board and says the run leaves out sealed items. The local server is POST /v1/systemone. The write-up is Strands Decider.

What is Clef-omni?

Clef-omni is Cloudflare's October 9, 2026 decision model, at $0.15 per million input tokens. The blog's BANKING77 cell is macro-F1 94.8. A 21-second video with sound scores in about 1.5 seconds. The weights are Apache 2.0, from Qwen3-Omni-30B-A3B-Instruct. The write-up is Clef-omni.

What is PurpleMIST?

PurpleMIST-Flash-1.0 scores 0.668 accuracy on 15,050 translated African-language decisions. DeepSeek-V4-Pro on that harness is 0.669, and KL is 0.214 against 1.125. The card says the checkpoint is Apache 2.0, about 7.9B parameters, and text only. A self-run of Decision Index edition 0.3 prints 46.7, which is a different test. The write-up is PurpleMIST.

What is vegaml?

vegaml scores 75.4% on 800 phishing emails against Jev 1.13.0 at 61.9%. The default checkpoint is 0.8B on frozen Qwen3.5-0.8B, Apache 2.0, with a 4B sibling. Decision 2.0's Vega, at index 56.5, is a different checkpoint. The write-up is vegaml.

What is d1-ultra?

Opened October 16, 2026, Artificial Analysis Decision Index prints d1-ultra at 65 and Jev 1.14 at 57. That day's Artificial Analysis thread prints an 85% non-hallucination rate on AA-Omniscience-Decision against 2% for hosted d1. Liquid's blog, posted October 15, prices input at $0.042 per million tokens, output free. Model id d1-ultra. POST https://api.liquid.ai/decisions/v1/systemone. This desk did not call the API. The write-up is d1-ultra.

What is Celeris-1 Decision?

Celeris-1 Decision, posted October 8, 2026, prints 80.0% on 3,210 jev-bench items at a 67 ms median, and $0.04 per million input tokens. Opened October 16, Artificial Analysis Decision Index lists it at 52. Model id celeris-1-decision. POST https://inference.celeris.ai/celeris-1-decision/v1/systemone. GET /v1/models with no key returned HTTP 401. This desk did not send a billed POST. The write-up is Celeris-1 Decision.

What is Span-02?

Opened October 16, 2026, the Span-02 card prints Decision Index 0.3 text index 63.62. The blog prints thinking off at 61.2 against Jev 1.13.0 at 58.0. Weights are a Gemma 4 26B-A4B fine-tune. Hosted GET of Respan's public models list prints input_cost 0.01. This desk did not run the weights. The write-up is Span-02.

What is kettei-fat?

Opened October 19, 2026, JevBench ranks kettei-fat first of 38 hosted systems at composite 77.8. Opened October 17, the API board named 39 systems at that same 77.8. ImageJevBench v0.3.0 capability is 75.9. Docs price kettei fat at $0.06 per million input tokens. POST https://api.sqwish.ai/v1/decide. GET /v1/models with no key returned HTTP 401. This desk did not send a billed POST. The write-up is kettei-fat.

What is basal-omni-1.0-ultra?

Opened October 19, 2026, JevBench ranks basal-omni-1.0-ultra first of 184 open systems on Capability at 82.5. Composite among open weights is 71.8, rank 3. Apache 2.0, Gemma 4 31B. The card prints Werdykt v1 at 0.879. This desk did not run the weights. The write-up is basal-omni.

Did HiddenLayer trick Jev on a checkout ad?

HiddenLayer's October 14, 2026 write-up prints TypeSafe Jev at 91% Confirm Price and 9% Place order on jev-ultrafast. Date headers 07:24:57 GMT and 07:24:58 GMT October 19, this desk's jev-1.13.0 still chose place_order, 1.0 then 0.89. Those two POSTs are not the 91 / 9. The write-up is HiddenLayer.

What is Drex Max?

Nace posted Drex Max on October 19, 2026. The product page prints Decision Index 0.3.2 public 67.59 against Jev 1.14 at 60.70. Docs price input at $0.06 per million tokens. Date headers 08:50:01 GMT and 08:50:02 GMT October 20, GET /v1/models with no key returned HTTP 401. The write-up is Drex Max.

What is Clairic?

Cloudflare posted Clairic on October 19, 2026, a System One extraction model on Workers AI. The blog's cascade ran in 19.3 s against 61.7 s for GLM-5.3 at every step. Docs price input at $0.10 per million tokens. It returns typed values, not a listed choice. The write-up is Clairic.

Did Datadog steer Jev with a counterfeit policy?

Datadog Security Labs' October 15 write-up, posted October 19, prints Jev Noul recall 82.90% on raw attacks and 3.88% with a counterfeit policy prepended. LangChain is_risky 40-run mean 0.893 then 0.290. Date headers 08:50:46 GMT and 08:50:47 GMT October 20, this desk's jev-1.13.0 is_risky was 0.94 then 0.13. The write-up is Datadog.

What is Sieve-31B?

Sthānika's Sieve-31B card, opened October 20, 2026, prints Decision Index 0.3.2 Full score 63.8. The authors' public-suite run is 62.52. Apache-2.0 LoRA on Gemma 4 31B. The write-up is Sieve-31B.

What is StartLux-Decision v1.1?

The StartLux-Decision README, opened October 20, 2026, prints v1.1 27B Instant at Full score 69.12, first of 174 on Decision Index 0.3.2. The October 3 self-run of edition 0.2.1 stays 63.88. The write-up is StartLux-Decision.

What is ToddlerJev?

ToddlerJev, posted October 18, 2026, prints 68.70% on MMAU-mini. Weights are TeenageWorks/Toddler-Jev under the Qwen Research License. This desk did not run them. The write-up is ToddlerJev.

What is pplx-decider-v1.2?

Perplexity posted pplx-decider-v1.2 on October 16, 2026. The card prints Decision Index 0.3 overall 68.09 on joint requests against v1.1 at 62.30. Docs price input at $0.02 per million tokens. Flash is $0.01. GET of the OpenRouter ids returned HTTP 404. The write-up is pplx-decider.

What is OpenRouter Armor?

On 492 tool calls, OpenRouter Armor's default Jev 1.13 scored 90.7%, 446 of 492, at $0.0021 a call. On October 16, 2026, at 01:18 UTC, this desk's jev-1.13.0 noul was 0.98 on git push origin main and 0.01 on ls ./reports. Armor is a beta plugin, not a catalog model. The write-up is Armor.

What is SubQ Snap?

On October 23, 2026, Date headers 02:23:17 GMT to 02:23:20 GMT, three POSTs from this desk to SubQ Snap returned HTTP 200. Docs name subq-snap-1.1 at $0.04 per million input tokens. GET /v1/models listed subq-snap-1.1 and subq-snap-preview. Artificial Analysis Decision Index prints Snap 1.1 at 65, tied with d1-ultra and Elephant 9B. The write-up is SubQ Snap.

What is mammoth-1?

Base Labs posted mammoth-1 on October 21, 2026, at 17:14 UTC. The card prints Decision Index 0.4 Full 71.24, first of 208. Opened October 23, the live space is edition 0.4.1, mammoth-1 Full 68.4 rank 3. Date headers 02:23:21 GMT to 02:23:23 GMT, GET /v1/models with no key returned HTTP 401. elephant-1 is the 9B sibling. The write-up is mammoth-1.

What is TrueForge with Jev?

Akshay Pachaar's weekly client follow-up agent inside TrueForge, TrueFoundry's MIT harness. On 40 client-issue matches, Jev and Opus both scored 37 of 40. Printed decision costs are $0.0015 with Jev and $0.005 to $0.015 without it. Review caught 5 wrong Jev matches. On October 13, at 09:21 UTC, he wrote that 40 is too small to establish how reliably confidence routes errors to review. On October 14, 2026, at 02:34 UTC, this desk's jev-1.13.0 call picked the intended issue at confidence 1.0. A trap shortlist scored 0.51. This desk did not install TrueForge. The write-up is TrueForge.

What is JevEval?

DeepEval's custom metric that sends bounded Noul, Score, and Choice questions to Jev in one request and returns a weighted mean. Apache-2.0, in deepeval 4.2.8. On October 14, 2026, at 01:58 UTC, two POSTs from this desk to jev-1.13.0 scored DeepEval's invented-breeze example at 0.254, and the matching clean answer at 0.901. The docs walkthrough that prints 0.461 is illustrative. This desk did not install deepeval. The write-up is JevEval.

What is Sage?

Opened October 11, 2026, the JevBench API board ranks Sage 1.3.0 first of 27 hosted systems at composite 74.0. Levanto's pricing page lists Pro input at $0.042 per million tokens and output at $10 per million for reasoning. The write-up is Sage.

What is H2O-Lightning-4B?

H2O's October 7, 2026 blog prints H2O-Lightning-4B v1.1 at JevBench composite 72.5, against Jev 1.13.0 at 71.5. That reading is board revision v1.7.17. The blog says the weights are Apache 2.0, from Qwen3.5-4B. The write-up is H2O-Lightning-4B.

What is OpenBles Firewall 1?

OpenBles Firewall 1 caught 119 of 120 agent attacks and flagged none of 120 benign look-alikes. The classic split is a different set, 55 of 60 against Jev at 33 of 60. The adapter is MIT, on fastino/GLiNER2.5-Decide. The write-up is OpenBles Firewall 1.

What is d3-ultra?

Opened October 26, 2026, Decision Index 0.4.1 Full score for d3-ultra is 70.18, second of 229, +3.03 over d3 at 67.15. The card's private SR-Text mean is 87.2. Apache-2.0, from Qwen3.8-27B. The write-up is d3-ultra.

What is d3-turbo?

Opened October 26, 2026, Decision Index 0.4.1 Full score for d3-turbo is 67.31, and vision full is 72.46, second of 56. The free API aliases vllm-sr/d3 to d3-turbo. Apache-2.0. The write-up is d3-turbo.

What is Decision Model Factory?

Yuma Ichikawa's arXiv 2610.34956 prints DF-JA-Large at 67.63 on S1MB Japanese Task Avg against a fixed public snapshot high of 60.05. A Hugging Face search on October 27 returned no matching public checkpoint. The write-up is Decision Model Factory.

What is hunch?

Matt Mastracci's October 27, 2026 posts name hunch as an EmbeddingGemma2 System One endpoint on CPU, at 20 ms on text and 100 ms on images. Cloudflare Clef teaches a local student in real time. Open-source launch was planned for later that week. Opened October 28, public search found no hunch GitHub repository and no hosted POST. The write-up is hunch.

What is Decision 3.0?

Opened October 26, 2026, Decision Index 0.4.1 Full score for d3 is 67.15. The October 16 card reading of edition 0.3.2 was 66.3. The free API aliases vllm-sr/d3 to d3-turbo. The write-up is Decision 3.0.

What is Mizan 4B?

xmanii/Mizan-4B prints a mean accuracy of 78.7 on 3,000 real Persian records, against the card's Jev cell of 83.7. The Q4 GGUF mean is 78.3. The card says the weights are Apache-2.0, from Qwen3.5-4B, and text only. The write-up is Mizan 4B.

What is Quyet-1.0-Large?

Quyet-1.0-Large's homepage prints frozen JevBench v1.6.0 at official score 71.39, rank 1 of 92, and Capability 81.7. The card says Apache-2.0, 31.3B, from Gemma 4 31B. A later reading on this site still lists Capability 81.7 and official composite rank 3. The write-up is Quyet.

What is Blink v0.3?

Pixilab's Blink v0.3 card prints a self-run of Decision Index 0.2.1 at 57.48, against the board's Jev cell of 57.91. The card's license field is Apache-2.0. The base is Gemma 4 26B-A4B. The write-up is Blink.

What is THX-01?

HAL-X's THX-01 card averages 98.4 on 2,843 support tickets, against Jev 1.13 at 97.4 on the same four sets. Number lookup on a different set is 93.4, and Jev on that line is 96.6. The October 11 post calls the API free. The docs page we opened does not print a price. The write-up is THX-01.

What is TypeLLM?

TypeLLM's September 23, 2026 file scores Qwen3.8-27B at 228 of 231 public JevBench tasks with thinking, and 195 of 231 without it. The hard tier moves from 76 of 111 to 109 of 111. Easy falls from 48 of 48 to 47 of 48. The homepage prices input at $0.05 per million tokens and thinking at $0.50. The hosted name is typellm-latest, and the 228 of 231 run is not that model. On October 16, at 05:44 UTC, TypeLLM posted that Jev 1.14 cannot return an array of vehicle boxes, including Tesla's flying Roadster. The write-up is TypeLLM.

What did You.com's risk sweeps cost?

Edward Irby's October 1 write-up prices three escalated sweeps at $0.43 in all, $0.14 each. You.com search is $0.37 of that total. Jev judgments are $0.0025. A blind read of nine Jev reports and six Qwen reports averages 4.25 on both sides. Qwen escalated 6 of 11 attempts. Jev escalated 9 of 9. The write-up is the risk sweeps.

Can Jev judge NSFW text?

jevmod scores sexual labels on OpenAI's moderation eval at AUROC 0.984, remeasured September 23, 2026. F1 at the shipped 0.80 line is 0.866. The question also names graphic gore. ZeroNSFWBot scores pictures on the host and asks Jev about text when a key is set. That repo prints no accuracy on adult images. The write-up is NSFW text.

What comes back from a call?

TypeSafe's primitives page lists three. Choice picks among up to 255 named options and returns a distribution. Score returns a number on a rubric you define. Noul is a yes/no and returns a probability between 0 and 1 without a separate confidence field. Several questions can share one request. TypeSafe says sampling is parallel and cites 70 to 500 ms. The model does not generate text. If the job needs a string, callers still use a language model.

Where can I send a request?

Access lists the public doors this desk has already cited: TypeSafe's API, Vercel AI Gateway, Cloudflare Workers AI, Netlify, OpenRouter, Venice, and LiteLLM. typesafe/jev-router is OpenRouter's chat model id. Arena's October 7 thread prints +3.6% net and a median $0.155 a task. The text comes back from the model that served the request. TypeSafe posted on September 20 that Jev is open with no waitlist, and that new accounts start with $5 of credit. Gateway quotas are not the same as a console key. Jevflake is a community dbt package that calls TypeSafe from Snowflake SQL. duckdb-jev is a native DuckDB extension over the same API. prompt_jev() is MotherDuck's hosted function on paid plans. Vercel's form router is a Next.js template on AI Gateway: Jev at confidence 0.95, Luna Fast below that. jsort is a command-line ranker over a TypeSafe key or OpenRouter. jevctl is a shell CLI over those same doors. n8n-nodes-jev-classification is a self-hosted community node with a Needs Review branch. On September 21 TypeSafe posted API and console issues, then We are back; the status incident is marked resolved at 08:16 UTC. On September 22 the same account paused new signups and said existing signups continue. On September 24 it said the signup banner was removed and that signups were still closed. On September 27 at 22:30 UTC it posted that signups were open again, and that new signups no longer get free credits. On September 30, 2026 we sent one POST to api.typesafe.ai/v1/systemone with model jev-latest. The state was "Help! My payouts have been failing for 3 days." The response was HTTP 200 and named jev-1.13.0. is_urgent, a Noul, returned 0.95. department, a Choice, selected billing, with probabilities 0.88 for billing, 0.12 for technical, and 0.0 for sales, and confidence 0.82. frustration, a Score, returned 1.04. Its legend is 0 Calm, 1 Frustrated, 2 Very angry, with probabilities 0.0, 0.96, and 0.04, and confidence 0.94. Usage was 426 input tokens and 73 output tokens. On October 6, 2026, at 17:05 UTC, this desk signed in to a new console account. Available credits were $0.00. Vercel, Cloudflare, Netlify, OpenRouter, Venice, and LiteLLM were not called.

What does TypeSafe publish as the price?

Opened October 7, 2026, the models page lists 80 requests per second for jev-1.13.0. Opened October 1, 2026, the models page lists jev-1.13.0 at $0.042 per million input tokens, free output, 100K tokens per second and 40 requests per second, 64k for the whole request and 32k for state plus the longest question. The page says those limits can change without notice. The price, the token rate, and the two context caps were the same on the October 7 reading. Opened October 15, 2026, those caps are still the defaults for Jev 1.14 and Jev 1.13. Date headers 08:49:11 GMT to 08:49:40 GMT October 26, POSTs with jev-latest and jev-preview named jev-1.14.0. Opened that hour, the models page says both aliases point at jev-1.14.0. On October 15, jev-preview pointed at jev-1.14.0 and jev-latest still pointed at jev-1.13.0. TypeSafe-Danger-Zone-Long-Context: true raises both caps to 262k. Vercel's model card listed input and output as Free, with promotional pricing printed as ending September 25, 2026. Opened again on September 26, the card still listed Free and still printed that date. Other hosts publish their own rates, which can go stale.

Who ran the speed and accuracy tests?

Evals collects the measurements already in the stories, with the caveat that came with each one. TypeSafe's workflow table is vendor-reported. Independent write-ups are author-run. JevBench is a 534-decision suite with a four-axis composite; Luna wins the hard items there and loses the score on price. DocJev put Jev and Luna on 40 PDFs and eight packets of the same LiteParse text. RZ's security demo put Jev next to Terra and Opus on 100 synthetic cases. jevsearch published Hit@1 on 41 labelled TypeSafe-docs queries. Open-Jev published correct counts on 231 public JevBench tasks: 9B 179, Jev 200, Luna 206, Astra 231. jsort published Pearson r 0.824 on 300 CommonLit excerpts. Pairwise Choice is this desk's eight-passage Stripe shortlist: a 0-9 Score put Roll secret at 8.86, and the other six scored 2.42 or below. JevEval is this desk's application of DeepEval's published equation to two hosted jev-1.13.0 calls, 0.254 and 0.901. TrueForge is Akshay's 40-request match, 37 of 40 with Jev and 37 of 40 without it, printed decision costs $0.0015 and $0.005 to $0.015. He wrote that 40 is too small for the confidence cut. MotherDuck published 89% on 100,000 AG News training rows. jev-gateway published token medians from 120 chess-agent sessions. AgentOS published a 121-turn routing chart, Jev at 0.901 against a local classifier at 0.355. laya-mlx published on-device latencies for Laya's weights, 13.42 ms and 7.39 ms on an M3 Max. Decider 1 published 0.768 against Jev at 0.727 on teacher-average labels. dejevu published a 5.63 second Flights run beside jev-ultrafast's published 7.09 seconds. A private banking feed published Jev at 36.1% on 612 rows. rev published public-tier rates, hard 0.784 for Qwen3.8-27B against its Jev cell at 0.730. hotdog-27B published 1,603/1,631 on a binary holdout. spark-s1-4b-v6 published a public-proxy Intelligence of 83.1. A Japanese table published JamC-QA at 56.0% for Jev and 76.5% for Luna. A 25-phrase chart published rank agreement 0.99 with 99 people. ASO Workbench published Spearman 0.74 on 578 terms. Jevsus published a September 20 diode: 209,628 calls, $3.13, and 45% of statements moving by 0.20 or more across three wordings. Rizzo Flow published 0.812 on SemIf's authored144, against SemIf at 0.819. Jevre published 0.9323 on its own test split. A computer-use README published $0.0002 against $0.032 on one screenshot. A 42-claim check published five runs: hard-18 means 73.3% and 70.0%, p = 1.000. A translation corpus published $0.3165 for 1,072 Jev scores. A local file search published nDCG@10 of 0.377 and 0.745 after JEV reranked. Jev RAG records TypeSafe's CLERC cookbook at top-1 5% to 18% on 40 queries, and Hindsight's LoCoMo recall@1 at 0.950 against 0.800. A needle script published 1,081 seconds against 1,480 on one file. CLM-8B published 31 of 38 on a DeepSWE verifier holdout. Eikos published 82.9 on its public hard row, against Jev at 73.0. A September 24 chess screenshot scores 388 games at 53.9% for Eikos. A fit table published 0 illegal moves in 489 MOBA calls. Opened October 8, 2026, the JevBench API board is release v1.6.1 under revision v1.7.21, ranks 26 hosted systems, and puts Jev 1.13.0 fourth at composite 71.5, with capability 77.1. The same board opened October 7, 2026, under revision v1.7.12, ranked Jev third at that score. JevBench v1.5.0, opened September 29, ranks Cygnet 73.7, Winnow-12B Q8 73.2, and Jev 72.1. The page calls the top pair a statistical tie. Jev's capability there is 80.0, first among 56 Jev-class systems. The older v1.4.2.2 cell is still 63.29, with imajev-4b at 67.37 on that board only. The v1.5.6 page opened October 4 lists the same checkpoint at 70.4, official rank 10. Imajev reads photos. Plumb prints 89 of 111 on its own public-hard runner. A 400M encoder named Kev prints 79.96% on 17,476 of its own rows. jevgrep prints 8 of 10 solves and a 25.8% lower combined bill. toolgate prints 19 of 20 on ten matched pairs. A grants thread prints 75 of 78 nos hidden. jevbench.app prints 20 of 21 on one recorded set. Decision Index 0.2 locks Jev at 51.67 in the kit's test. The September 28 board file is edition 0.2.1 and lists Jev at 57.91. Opened October 8, 2026, the space still headers v0.3, 115 models, updated 2026-10-07, and the head-to-head panel still ranks Perplexity Decider v1.1 first at 62.8, with Jev 1.13.0 as a reference at 60.1. Opened October 7, 2026, the header said 112 models, updated 2026-10-06. Drex DLM prints 52.31 on a card labelled edition 0.2, against the kit's Jev cell of 51.67. PolicyLM-1.7B prints 0.842 at a 0.5 cutoff, and that table has no Jev column. GLiDE v1.2 prints Decision Index 0.4.1 Full 71.5, rank 1 of 229, opened October 26. The October 1 self-run of 64.81 stays. Jeva-4B prints 0 flips on 600 Choice items. BANKING77 at K=16 is 91.62 against Jev 82.50. Abe's 1,774 posts cost $0.055 on Jev Noul against $1.87 on AI_FILTER. GEV-26B-Decide prints 62.48 on another 0.2.1 self-run. 64.81 and 62.48 are not rows in the September 28 file. pplx-decider-v1.2 prints Decision Index 0.3 overall 68.09 on joint requests against v1.1 at 62.30. The October 1 panel still prints 85.71% for v1. kettei-fat is first of 38 hosted systems on the JevBench API board opened October 17, at composite 77.8. StreamDecisionBench prints Kev-27B at 62.15 and the Jev API at 59.63. A September 29 abstract counts 346,009 requests for under $10. The page we opened September 30 is JevBench v1.5.4, with 106 ranked systems. The September 29 addenda are numbered on that page and still tagged roster addendum. Matilda prints a self-run score of 59.26. That 59.26 is not in the September 28 board file. OpenAI's Decisions API is a separate call on OpenAI's own API. TechCrunch on September 30 quotes Sam Altman on speed and says TypeSafe did not answer its questions. OpenAI Jev records Date headers 02:23:12 GMT and 02:23:13 GMT October 23, two POSTs HTTP 200. OpenAI Developers posted general availability on October 22. The guide prices uncached input at $0.10 and cached input at $0.01. The October 6 public-beta POST also returned HTTP 200. Clef is Cloudflare's October 1 pair. The October 1 blog median is 209.3 ms and 38.8 ms against a printed Jev cell of 524.1 ms. The October 1 cards list $0.24 and $0.09 per million input tokens. The October 9 blog lists Clef-flash at $0.038. Clef-omni is $0.15, and its BANKING77 cell is macro-F1 94.8. PurpleMIST-Flash-1.0 scores 0.668 on 15,050 translated African-language decisions. vegaml scores 75.4% on 800 phishing emails against Jev 1.13.0 at 61.9%. Sage 1.3.0 is third at composite 74.0 on the JevBench API board opened October 17. Opened October 11 it was first of 27 at 74.0. H2O-Lightning-4B v1.1 prints composite 72.5 against Jev at 71.5 on the October 7 blog. OpenBles Firewall 1 caught 119 of 120 agent attacks. d3 prints 66.3 on Decision Index 0.3.2, opened October 16. The October 11 reading was 64.7 on edition 0.3.1. Mizan 4B averages 78.7 on 3,000 Persian records. THX-01 averages 98.4 on 2,843 support tickets. Quyet-1.0-Large prints Capability 81.7 on frozen JevBench v1.6.0. Blink v0.3 prints 57.48 on a self-run of Decision Index 0.2.1. d1-ultra is the only decision model Vals AI printed with all 1,960 ClaimBench claims right. Opened October 16, Artificial Analysis Decision Index prints 65 against Jev 1.14 at 57. mammoth-1 prints Decision Index 0.4 Full 71.24 on the card. Opened October 23, live 0.4.1 Full 68.4 rank 3. Celeris-1 Decision prints 80.0% on 3,210 jev-bench items at a 67 ms median. That board lists it at 52. Span-02 prints Decision Index 0.3 thinking off at 61.2 against Jev 1.13.0 at 58.0. OpenRouter Armor scored Jev 1.13 at 90.7% on 492 tool calls, 446 of 492. SubQ Snap 1.1 prints 65 on Artificial Analysis Decision Index. Docs name subq-snap-1.1 at $0.04 per million input tokens. Date headers 02:23:17 GMT to 02:23:20 GMT October 23, three POSTs returned HTTP 200. 61.7 is the October 12 company Decision Index 0.2.1 cell. JevEval scored 0.254 on this desk's invented-breeze call to jev-1.13.0, and 0.901 on the clean answer. TrueForge scored 37 of 40 on Akshay's client-issue match, the same count as Opus, at $0.0015 against $0.005 to $0.015. He wrote that 40 is too small for the confidence cut. Liquid d1 prints 58.9 against Jev at 57.9 on an internal Decision Index chart. Open d1 prints d1-3B at 48.57 on a self-run of Decision Index 0.2.1. Winnow-12B on that table is 50.02. The table has no Jev row. Vela 2.0's 9B card prints 41.63 with Noul calibration on, and 41.09 with it off. The evaluation file prints Lux at 46.23 and has no Jev row. Step-Jev scores 19.6% when the reward grades every search step, and 5.4% when it grades only the final answer. On October 9 the repository's no-step test state, and a no-search answer of 1965 written here, both scored incorrect at 1.0 on jev-1.13.0. The rollout files for real_jev_gate.py are not in the repository. A bag of Jev questions moves a published WANDS mean from 0.574 to 0.610. The median stays 0.561, and the CSV stores one float for both rows. Pairwise Choice on eight Stripe passages this desk wrote put Roll secret at 8.86 on a 0-9 Score. Pairwise ranked rotating the API key third, and the majority graph had a cycle. Drex 1.5 prints 199 of 231 against Jev at 201 of 231. The post calls it first on JevBench. Rox prints Jev-Score at 96% and a production reranker at 84%, at 200 chunks, mean over 75 queries. Jev Sentinel's demo page prints about $55 per million actions. TechCrunch prints $2.94 against $372 for one job, and those two figures are not on the demo page. Mercury Decide cites JevBench v1.4 and prints no score. Opened October 8, 2026, the API board ranks it third at 72.4, and the free model page returned, priced Free. An October 1 search of the live JevBench page found no Mercury Decide and no Drex 1. Microsoft-Decision-1 says it was 4.5 times quicker than Quyet-1.0-Large on Microsoft's own comparison, and does not print an accuracy. Opened October 10, 2026, the JevBench API page still said release v1.6.1 under board v1.7.29, and a search of that text found no Microsoft-Decision-1. TypeLLM scores 228 of 231 on its own public-subset file with thinking, and 195 of 231 without it. That 231 is not Drex's 199 of 231. TypeSafe's Deel charts print 6 of 8 quality checks at parity or better, and a 16.6-point drop on messy metric questions. A registered attack paper prints validated success on jev-1.13.0 from 1.8% to 3.5%. RLCDAlignBench prints a median AUROC of 0.886, and a median ECE of 0.168. jevos-1b prints 0.815 against Jev at 0.927 on 2,000 yes/no questions. jev-code prints three coding tasks that still fail. astra-jev-harness prints a 27% drop in Astra input tokens on one two-file fixture. Strands Decider 2B prints 167 of 231 on the public tasks, median 115 ms on an RTX 3090. HydroJEV prints mean macro-F1 0.63 in distribution, and a gate on 69 of 200 and 76 of 200 windows. A SemEval table prints Task 1 RMSE 1.0645. jev-doc-search prints two filing questions, both marked correct. We did not rerun them. Compare puts similar tests next to each other. Limits is where those tests, and TypeSafe's own list, say the model fails.

What is TypeSafe's rule on publishing benchmarks?

The Master Customer Agreement section 2.3(f) forbids customers from publishing benchmarks or performance information about the Services. Several stories report numbers people posted anyway. We did not ask TypeSafe whether those posts are permitted. See method.

Why do open decision models start from an open language model?

A closed API does not release the weights a decision head has to read, so the published checkpoints start from a public language model. Jeff v1.3 prints a mean of 30.4% for untrained Qwen3.5-0.8B and 92.8% with an adapter, across 13 of its own tests. Unsloth's October 7 documentation also fits Llama 3.2 3B and Gemma 4 E4B. The write-up is Open decision models.

Are Jev's weights public?

Jev's weights are not public. SemIf, Laya, Nimble, jeff, LocalJev, Verdict, Kev, djev, SimpleJev, llamacpp-jev, rev, spark-s1, hotdog-27B, Rizzo Flow, Jevre, jev_local, decider-4b v2, Imajev-4B, Plumb-4B, a 400M encoder also named Kev, and jevos-1b are community stand-ins with their own evals. kishida/llama.cpp branch jev builds the same path into llama-server, and its table puts Qwen3.8 Flash-Next at 0.944 on 1,191 questions. On October 2, llama.cpp itself merged /v1/systemone for Julia-1, Laya, Kev-4B, lev, and OpenJev. Release v0.6.0, published October 5, also serves Nimble and Clef. The server README at that tag names Clef as an image model. Opened October 6, the kishida page still says the endpoint is not upstream. jevos-1b prints 0.815 against Jev 0.927 on 2,000 yes/no questions in the README image alt. Choice and Score return 422. Imajev-4B scored 67.37 on the v1.4.2.2 harmonic mean. On v1.5 that row is a roster addendum at 70.4, outside the ranked order. Cygnet leads v1.5 at 73.7. Jev is 72.1. The 400M encoder is not Palmer's Kev. laya-mlx is an MLX port of Laya's published checkpoints. SemIf's Qwen3.5-4B row on JevBench is 74.7 against Jev 75.4. djev is 74.3. DiffusionGemma as Jev is the one-step canvas read. A hand chart adds to 198 of 201 against Jev at 191 of 201. Laya is 70.1. SimpleJev Qwen3.8-27B is 67.3. Open-Jev's released 9B checkpoint is 179/231 on the public JevBench tasks, and the 27B v1.1 card is 197/231 with a hard tier of 80/111, against Jev at 200/231 and 81/111 on that same slice. rev's frozen 27B note is hard 0.784 on 111 public items. spark-s1-4b-v6's public-proxy Intelligence is 83.1. Rizzo Flow's Q8_0 note is 0.812 on SemIf's authored144, where SemIf published 0.819, and 6 of 36 missing-evidence rows stay confident. Jevre's test split is 0.9323 with no Jev column, and held-out sources are 0.8366. None is a TypeSafe weight release. Strands Decider 2B is an October 1 Apache 2.0 release, 1.9 billion parameters, 167 of 231 on the public tasks. Bud Decision Studio serves eleven of those open checkpoints on a local TypeSafe-shaped API at port 8420. DSPy 3.4.0 calls hosted Jev from Predict. The patterns repo pins jev-1.13.0. Repos groups those projects.

What is this site?

Jev News records decision models that return a choice, a score, or a yes/no from a fixed answer set. Stories are rewritten from public posts on X and the pages those posts cite. Original URLs sit at the bottom of each article. This site is not affiliated with TypeSafe AI. About and method are the longer version of those rules. Demos and waitlist notes stay in the news archive.