you can make any open source model behave like jev with just a bit of inference engineering. it's shockingly easy. to prove it, we built a new endpoint we're calling deepseek-v4.1-flash-jev. see the demo below. here's how it's done: sglang (an inference engine) offers a Show more
CommunityPractices & Patterns
Can LLMs Do What Jev Does? A Replication Attempt
A skeptical Japanese replication that approximates Jev's speed by reading first-token logits from small LLMs, with Gemma3 and a Mario benchmark.
- Category
- Practices & Patterns
- Format
- Article
- Published by
- Community
- Author
- nwn
- Added
- 2026-09-25
- Last verified
- 2026-09-25
Highlights
- Reimplemented parallel decisions with Gemma3 270M and first-token logits, reporting a 77x speedup.
- Mario harness: Jev 276 ms and 1,226 distance against Llama 3.1 8B at 356 ms and 1,138.
- Notes only non-reasoning models expose logprobs for this approach.
- A commenter's calibration test saw Choice pick one option in all 400 hidden-die rolls at 83% average.
Watch out
Japanese-language with small samples, a reverse-engineered prompt, and older logprob-exposing models only.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
Hacker News: Jev implemented in 25 lines of Python
bashbjorn400 points · 128 commentsJev in 25 Lines of PythonRead the thread on Hacker NewsHacker News: Kev: a tiny Jev-like family built on Qwen3.5
tosh458 points · 199 commentsKev: Tiny Jev-like family of decision models built on top of Qwen3.5Read the thread on Hacker NewsX: deepseek-v4.1-flash-jev
Reddit: A walkthrough of routing between models with Jev.
More like this
4.3kGitHub stars
An independent baseline that reads typed option probabilities straight from a frozen Qwen3.5-4B's logits in one forward pass, reproducing Jev's interface pattern with open models rather than Jev's undisclosed model or training.
Practices & Patterns#community#python#open-models
Community
TheoLeeCJ/SemIfPython
TheoLeeCJ/SemIfPython6.9kGitHub stars
Apache-licensed, locally runnable Jev-style decision models (0.8B, 4B, 9B on Qwen3.5) with released weights, training code, a System One-compatible server, frozen eval suites, and a playground.
Practices & Patterns#community#python#open-models
Community
jaredpalmer/kevPython
jaredpalmer/kevPython404GitHub stars
An independent open reproduction of the System One model class: Qwen3.5-based 2B and 35B mixture-of-experts models that return typed Choice, Score, and Noul probabilities in one forward pass, with nothing distilled from Jev.
Practices & Patterns#community#python#open-models
Community
Mapika/deciderPython
Mapika/deciderPythonFrom the community
Posts from builders shipping with Jev right now.
Tool calling as classify plus action, back to twelve-factor agents
jev is the best excuse you could possibly have to go re-read 12 factor agents. Tool calling itself can be decomposed into classify+action, if you learn to design ai programs as pipelines that switch breathlessly between classification, structuring data, deterministic code, AND Show more
@dillon_mulroy
i think jev is resonating with devs so well b/c it unlocks so many opportunities for composing ai into systems and products rather than ai _becoming_ the product/system really does feel like it was a missing primitive
Why Jev encodes a state and outputs distributions instead
Jev is one of the more interesting model launches I have seen recently because it asks a very simple question: Why are we using autoregressive LLMs as insanely expensive if statements inside software? The easiest way to think about it is: LLM: text -> generate tokens Show more
@CompleteSkeptic
After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI? I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev • 20-200x faster • 40-400x
How the agent stack moved in Jev's first three days
Jev just landed and the agent stack moved in three days. TypeSafe AI released Jev on September 15. It is built for the work agents actually do most of the time: choose the next tool, score a risk, decide whether to retry, or answer yes or no. Latency is under half a second. Show more
Jev plays Doom at ~10 calls/sec for ~$7/hour
We love how this doomo doomonstrates real-time intelligence and what can be doone with code + AI! ~10 calls/sec = ~$7/hour
Jev plays Super Mario Bros.
got @typesafeai’s new model Jev to play Super Mario Bros. fast inference + structured outputs makes it surprisingly good for real time use cases. I'm excited to see what can be done with these new models!
LLMs generate answers, Jev makes decisions
this is the easiest way to understand Jev: LLMs generate answers. Jev makes decisions. that sounds like a small difference, but it actually changes the entire use case. say you give a normal LLM this: “here’s a user, their account history, payment behavior, support chats, Show more
@CompleteSkeptic
After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI? I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev • 20-200x faster • 40-400x