Jev, the model that refuses to write, draws Silicon Valley crowds and a $40m seed round

Update

What changed

The first outside numbers on Jev have arrived, and they largely back the sales pitch. Pranit Sharma, an engineer at Vercel, says his team had been using ChatGPT Luna 5.6 to check commands for safety. After switching to Jev, responses arrived 5 to 18 times faster, with classification accuracy improved into the bargain.

Bryo AI chief technology officer Nikhil Mudholkar ran a different race, pitting Jev against Gemini on classifying business correspondence. His verdict: Gemini edged it on accuracy but cost 10 to 20 times more, while Jev returned something the alternatives did not. 'This is the only system that returns a true probability, which makes it ideally suited to workflow automation,' he said.

That probability is the load-bearing part. Armin Ronacher, CTO of Earendil, argues the score lets responsibility be divided sensibly: a 50% confidence answer can be bounced back to a person, while 95% may be grounds for the system to act on its own. The economics lean the same way, with no output tokens billed at all and input measured in billions rather than millions of tokens. TypeSafe also pitches Jev as a watchdog, analysing agent activity logs, flagging dangerous behaviour, blocking attempts to slip past restrictions, and routing tasks to whichever model should handle them.

Demand, at least, needs no benchmark.

Sources and evidence

• TypeSafe AI Unveils Jev, a Fast AI Model Built for Software Automation | Ukraine news – Mezha – Межа. Новини України.: TypeSafe says its decision-only model Jev, which returns calibrated probabilities over predefined options instead of generating text, briefly could not serve all API users at launch; named early users report 5-18x faster command-safety checks than ChatGPT Luna 5.6 with improved accuracy, and business-correspondence classification slightly below Gemini's accuracy at 10-20x lower cost; TypeSafe says Jev is trained exclusively on synthetic data using reinforcement learning based on calibrated decis

Independent WittyWires Watcher; not an official account or feed.

Update

What changed

Jev is collecting an ecosystem at speed. Per Pat McGuinness, writing on his AI Changes Everything newsletter this evening, Cua has shipped jev-use, a development preview that pairs the model with Cua Driver so low-latency decisions handle the clicking, typing and scrolling of computer use while the surrounding agent gets on with the reasoning. Alex Volkov has released Jevify, an open-source skill that scans an application or codebase for decision tasks Jev could take over, then builds tests comparing latency, cost and quality against whatever runs there today.

The plumbing is spreading too. He reports the model is now available across developer environments, with direct integration support on Cloudflare's serverless AI platform and early agent harness integrations. For a launch measured in days, that is quick work.

The sceptics have organised as well. TypeSafe has published a blog post describing its RLCD training method but no technical paper, and outside observers suspect the architecture rests on an open-weight LLM. The sharper datapoint is Laya, an open-source classifier with 421 million parameters on a ModernBERT backbone, which he reports performs comparably using the same three output types: a choice from a list, a boolean, a score.

Sources and evidence

• Jev Makes Fast and Cheap Decisions – Substack: Days after launch, TypeSafe's Jev has attracted working ecosystem tooling (Cua's jev-use computer-use preview, Alex Volkov's Jevify auditing skill, Cloudflare serverless availability) and organised reproduction scrutiny, including a community tracker of more than a dozen efforts and a 421M open-source classifier, Laya, reportedly performing comparably, per Pat McGuinness's reporting.

Independent WittyWires Watcher; not an official account or feed.

Update

What changed

The Jev story has a useful new layer: SSBCrack identifies TypeSafe AI's founder as former OpenAI researcher Diogo Almeida and reports that Jev is built to return calibrated probability decisions instead of generated text.

The article also attributes early results to users at Vercel and Bryo AI, including five-to-eighteen-times faster safety-command classification, improved accuracy and confidence scores designed for workflow automation. These are reported claims, not independent benchmarks.

Almeida also sees Jev working as a routing and monitoring layer around larger language models, which sharpens the architectural pitch: use a small decision model for the countless yes/no-ish turns, and keep the heavyweight model for work that actually requires synthesis.

Sources and evidence

• SSBCrack reports.: TypeSafe AI, founded by former OpenAI researcher Diogo Almeida, has launched Jev, a transformer-based model designed to output calibrated probability decisions rather than generated text.

Independent WittyWires Watcher; not an official account or feed.

Update

What changed

Jev’s pitch is becoming clearer: it is not a smaller chatbot trying to write better prose, but a transformer-based model designed to return predefined probability decisions instead. Digital Today, relaying TypeSafe AI and TechCrunch, says developers are testing it for software automation, with output tokens free and input charged in billion-token units rather than million-token units.

TypeSafe also says demand briefly made API access difficult. That sharpens the practical case for Jev as a low-cost decision layer around larger models, while leaving the important questions open: how calibrated are those probabilities, and how well do they hold up outside early adopters?

Sources and evidence

• Digital Today reports TypeSafe AI's description of Jev.: TypeSafe AI's Jev returns probability values or predefined decisions instead of generating text.

Independent WittyWires Watcher; not an official account or feed.

Update

What changed

TypeSafe’s Jev launch now comes with enough practical detail to move beyond the initial pitch. MarkTechPost reports that the hosted service is in early access behind a waitlist, with a single API endpoint and SDKs for Python and JavaScript. Developers send it a state and typed questions, rather than asking for a block of generated prose.

The service supports Choice, Score and Noul question types, with questions evaluated in parallel against the same state. Its confidence values are intended to support a three-way workflow: act when confidence is high, review the middle ground and send low-confidence decisions to a person. TypeSafe says Jev costs $42 per billion input tokens, with output tokens free, although the launch figures and comparison tests come from TypeSafe’s own workflow evaluations.

The launch also gives Jev an early ecosystem. MarkTechPost lists integrations and experiments covering command safety, email triage, browser and phone agents, database filters, smart-home controls and agent guardrails. Those examples show where the model is meant to sit, as a fast decision layer inside larger systems, but they are demonstrations and attributed claims rather than independent proof of calibration or reliability. The useful new question is no longer whether Jev can return a probability.

Sources and evidence

• TypeSafe AI Releases Jev: A System One Model That Returns Typed, Calibrated Decisions Instead of Text – MarkTechPost: TypeSafe has launched Jev in early access as a hosted, transformer-based system that returns typed decisions and probabilities rather than text, with developer tooling and early agent integrations.

Independent WittyWires Watcher; not an official account or feed.

Update

What changed

Developer Teknium says a run of Jev's public compaction evaluation produced a programmatic rule rather than a useful model-led decision. In his assessment, the result did not need Jev or another model to produce the outcome.

Teknium describes that as the central problem with the approach: the compaction process, he says, effectively stripped away the need for the model instead of demonstrating a valuable decision shortcut. That is a sharper criticism than a merely disappointing score, because it questions whether the system is adding intelligence at all.

This remains one practitioner's account of a public evaluation, not an independent benchmark or a finding that applies to every Jev workflow. The useful next test is whether other developers can reproduce the result, and whether TypeSafe can show cases where compaction preserves a meaningful decision that a simple rule could not make.

Sources and evidence

• Teknium on X: Okay, everyone wants us to give the unbiased facts. Jev compaction here has a lot of problems. The biggest problems: – When I ran it on our public compaction eval (th: Teknium says Jev compaction produced a programmatic rule that did not require Jev or another model in his run of the public compaction evaluation.

Independent WittyWires Watcher; not an official account or feed.

Update

What changed

Update, 18 September: the founder, the fans and the fine print

The launch story now has a named author. TechCrunch reports that TypeSafe was founded two years ago by Diogo Almeida, a former OpenAI researcher who worked on ChatGPT and on the reinforcement learning from human feedback method that helped shape how modern assistants behave. For a company asking developers to hand a new kind of model the decisions inside their pipelines, that pedigree is the argument.

TypeSafe's own account of the launch adds the claims worth testing. Because users define Jev's possible outputs in advance, the company says the model never generates text and therefore cannot hallucinate in the sense that keeps language-model users up at night. The economics run backwards too: output tokens are free, and input is measured in billions rather than millions. Demand on launch day was strong enough that the API briefly could not serve everyone, which is either a triumph or an unplanned load test, depending on your temperament.

The named adopters have started doing the arithmetic in public. Vercel engineer Pranit Sharma says his company replaced a ChatGPT Luna 5.6-based classifier that checked command security, and that Jev returned results 5 to 18 times faster with higher accuracy.

Sources and evidence

• TypeSafe unveils Jev model for software automation — TechCrunch – UA.NEWS: TechCrunch-sourced reporting says TypeSafe's decision-only model Jev was founded two years ago by ex-OpenAI researcher Diogo Almeida, trains exclusively on synthetic data via reinforcement learning from calibrated decisions, and is being adopted for classification and model-routing work, with Vercel reporting 5 to 18 times faster command-security checks than its Luna 5.6-based setup and Bryo AI reporting a 10 to 20 times cost advantage over Gemini at slightly lower accuracy.

Independent WittyWires Watcher; not an official account or feed.

Update

What changed

TypeSafe’s Jev now has an open-source challenger called Laya, released four days after Jev’s launch. The new model returns typed decisions rather than generated prose, putting fast, local AI classification within reach of developers with a single T4 GPU rather than a hosted API account.

Laya ships three Apache-2.0 checkpoints: a 421-million-parameter English model, a 322-million-parameter multilingual model covering more than 100 languages, and a typed-decisions model. Its creator reports 32.8 milliseconds per question on a T4 GPU and 7.2 milliseconds in batches, compared with reported Jev latency of 236 to 276 milliseconds. The model can run in a free Google Colab session, with no API bill.

The performance picture is not a clean victory lap. On a shared typed-decisions set, Laya’s reported score was 0.766 against Jev’s published 0.727, although those figures came from different benchmark runs. An independent tester using 78 cases found Jev scored 0.974, while Laya scored 0.590. Laya was still the fastest system in that test, at about 30 milliseconds per case against roughly 302 milliseconds for Jev. Small samples do not settle the argument, but they do make the word “better” work for its supper.

Laya’s release also brings a dispute over credit.

Sources and evidence

• TypeSafe Jev: System One Model, answered by Laya – Gadget Pilipinas: An open-source model named Laya has been released as a challenger to TypeSafe’s Jev, offering local typed-decision inference with reported millisecond-level latency, while available testing presents a mixed accuracy picture and a researcher disputes TypeSafe’s originality claim.

Independent WittyWires Watcher; not an official account or feed.

Update

What changed

TypeSafe AI opened early access to Jev on 15 September after two years in stealth and a reported $40 million seed round led by DCVC, according to StartupFortune citing Dealroom. The model is designed to return typed choices, scores and probabilities rather than generated prose, positioning it as a decision layer beneath larger AI agents.

The early adoption figures are striking but remain reported claims. StartupFortune says Vercel described Jev as the fastest-adopted model in its AI Gateway history, with nearly 13% of paid teams using it by the first 24 hours. The same account says Jev is priced at $0.042 per million input tokens, with output free, and that TypeSafe’s own tests found major speed and cost advantages on narrow decision tasks.

That gives the launch a more practical shape. Jev is aimed at the small judgements inside software, such as classifying a request, checking a tool call or routing a ticket, where paying a general-purpose language model to write a tiny answer may be overkill. Vercel and Cloudflare integrations suggest it is already being placed where developers build and run applications, although the adoption, benchmark and reliability claims have not been independently tested here. The real question is whether its probabilities remain trustworthy once messy production traffic arrives, rather than merely looking tidy in a launch window.

Sources and evidence

• TypeSafe AI's Decision Model Jev Becomes Vercel's Fastest Adopted Launch – startupfortune.com: TypeSafe AI’s Jev has moved from launch curiosity towards practical developer infrastructure, with reported early adoption through Vercel and integrations including Cloudflare, but its strongest performance and reliability claims remain vendor or publication-reported.

Independent WittyWires Watcher; not an official account or feed.

Update

What changed

Jev is attracting a burst of practical experiments from AI developers, who are using the model to make the small, repeated judgements that keep agents moving. A 36Kr report says developers have connected it to browser agents, Claude Code context-cleaning tools, advertising analysis and task-acceptance checks, while LangChain has released a Jev-as-a-Judge experiment for evaluating agents.

The model returns structured judgements rather than prose, using three reported formats: Noul for yes-or-no decisions, Choice for selecting from fixed options and Score for rating against set criteria. Multiple questions can be assessed against the same input, allowing a system to ask whether a task is complete, whether human intervention is needed or which action a browser should take.

The most eye-catching figures remain early evidence, not settled performance. 36Kr reports that a LangChain test put Jev at about 0.44 seconds per call and roughly $0.00035 per call, while an advertising experiment attributed to developer Matthew Berman processed 724 adverts and 8,724 judgements in about 40 seconds for around nine cents. LangChain described its experiment as small and called for testing on more agents and real production tasks.

Sources and evidence

• Why Is Jev Viral? Former OpenAI Researcher's AI Tool Taking Over The Internet — Why Everyone Is Scrambling To Use It – 36Kr: 36Kr reports that Jev is being tested as a structured decision layer for browser agents, coding workflows, advertising analysis and agent evaluation, with early experiments reporting low latency and cost but limited validation.

Independent WittyWires Watcher; not an official account or feed.

Update

What changed

TypeSafe AI’s Jev is designed to make structured decisions in roughly a tenth of a second, rather than generate prose. TechSpot reports that the model accepts a situation and predefined questions, then returns choices, scores or yes/no-style answers with probabilities, making it a potential decision layer for fraud checks, support routing and agent safeguards.

The practical pitch is speed and cost. TypeSafe claims response times of 70 to 500 milliseconds, with input priced at $0.042 per million tokens and no output-token charge because Jev does not produce a paragraph. In one illustration, an invoice decision that might take a conventional language model seconds is reduced to three figures showing fraud, clean and review probabilities. The model evaluates questions in parallel rather than writing an answer word by word.

There is a sizeable asterisk, and it is doing honest work. TechSpot says TypeSafe’s headline performance figures come from internal workflow tests, with Jev reaching about 67.8% agreement against averaged answers from other language models, compared with 74.1% for GPT-5.6 Sol. The company has not published an architecture paper or parameter count, and Jev’s current early-access version handles text and structured data, not images. A probability score can help decide when to involve a person, but it does not make a wrong decision right.

Sources and evidence

• Meet Jev: A New AI Model From One of ChatGPT's Co-Creators – TechSpot: TechSpot reports that TypeSafe AI’s Jev is an early-access AI model that returns predefined decisions and probabilities in 70 to 500 milliseconds, with low input pricing and free output, while its published internal tests show lower agreement than a larger comparison model and leave calibration and architecture questions unresolved.

Independent WittyWires Watcher; not an official account or feed.

Update

What changed

TypeSafe’s Jev is now being framed less as a faster chatbot and more as a decision layer for software. A new Hugging Face Community Article sets out when developers should consider it instead of a generative large language model, with the key dividing line being whether the output must return to application code as a choice, score or yes/no judgement rather than prose.

The article says Jev is suited to finite answer spaces, routing, ranking, blocking and escalation. Its stated interface returns typed decisions, probabilities and confidence signals, while multiple questions can be evaluated against the same piece of state. That could let an application route a support request, choose a model or send a borderline case to a person without parsing a paragraph of generated text.

The guidance is notably less breathless than a simple replacement pitch. It says generative models remain the better fit for writing, explanation, summarisation, open-ended research and creative work. It also says Jev does not replace business rules, permissions or audits, and that its probabilities still need testing against representative, ambiguous and adversarial inputs.

The practical recommendation is to measure the whole workflow, including latency, cost, retries, parse failures, human intervention and final task success.

Sources and evidence

• Jev AI vs LLMs: When Should You Use a Decision Model Instead of a Chat Model?: A new Hugging Face Community Article argues that Jev should be evaluated as a typed decision layer for finite, code-driven judgements, while generative LLMs remain better suited to open-ended content and reasoning. It also recommends testing latency, cost, error rates and human-review outcomes rather than relying on confidence scores alone.

Independent WittyWires Watcher; not an official account or feed.

Update

What changed

TypeSafe’s Jev now has a detailed practical guide showing developers how to use its structured decision model inside software, rather than treating it as a chatbot with better manners. The guide sets out a four-step flow: send a state, define typed questions, read the structured response, then let application code decide what happens next.

The model offers three question types. Choice selects from a defined set, Score rates something against an ordered rubric, and Noul judges whether a statement is true. A single state can support several questions in parallel, such as routing a support ticket, scoring its urgency and checking whether a refund is explicitly requested.

The useful boundary is that Jev does not own the policy. The guide says developers still set thresholds, permissions, audit logs and human-review rules. A high-confidence result might route a ticket automatically, while an uncertain or high-risk result can trigger clarification, a stronger model or a person. It also says Jev currently accepts text, JSON objects and arrays of text, not images, audio or video directly.

That makes the pitch more concrete, but not magically proven. The guide warns that probability is not a business-accuracy guarantee and recommends narrow questions, preprocessing non-text inputs and validating latency in the intended environment.

Sources and evidence

• What Is Jev AI? A Practical Guide to System One and Executable Decisions: TypeSafe’s Jev guide explains how developers can send shared state and typed questions to receive structured choices, scores or yes/no judgements, while keeping thresholds, permissions and final actions in application code.

Independent WittyWires Watcher; not an official account or feed.

Update

What changed

TypeSafe AI’s Jev has a more consequential origin story than its unusual output format alone suggests. ETV Bharat reports that the model was launched on 15 September by Diogo Almeida, a former OpenAI researcher who helped build ChatGPT and co-created reinforcement learning from human feedback, or RLHF.

Almeida left OpenAI after concluding that increasingly capable chat models were not proving especially effective at automation tasks. ETV Bharat says he spent two years building TypeSafe AI in stealth and developed a training approach called Reinforcement Learning for Calibrated Decisions, or RLCD. Almeida’s stated question was why superhuman chat models had not produced artificial general intelligence, a useful clue to the problem Jev is attempting to tackle rather than another slogan about making chatbots bigger.

The new report also adds a practical description of Jev’s intended role. Instead of generating prose, it takes a state and a typed question, then returns a predefined decision as a Noul yes-or-no percentage, a Choice or a Score. TypeSafe says confidence is meant to guide the hand-off: high-confidence decisions can be automated, while uncertain cases go to a person.

Sources and evidence

• Ex-OpenAI Researcher Launches Jev, A Chatbot That Makes Decisions Instead Of Text – ETV Bharat: ETV Bharat reports that former OpenAI researcher Diogo Almeida launched TypeSafe AI’s Jev on 15 September after two years developing a calibrated-decision approach for software automation.

Independent WittyWires Watcher; not an official account or feed.