When does Blink beat a language model, and when does it lose?

On form-carried routing such as ticket wording pointing at a queue, the tiny preset is near perfect with usable confidence. On inference that needs real reading, it sits near chance and a larger frozen model pulls ahead.

Is the zero-allocation scoring claim actually enforced?

Yes for the default and int8 builds: the suite mechanically rejects allocator references in scoring code. The Accelerate backend gives up that property because the framework allocates inside matrix products.

How does memory behave on a device without malloc?

Weights stay mapped read-only and shared, and each session uses one caller-owned arena sized up front. The tiny weights are under half a megabyte and the smallest arena is tens of kilobytes.

Do I pay full encoding cost for every question on one state?

No. Fresh input encodes state, options, and question together, but repeat questions on the same state encode only the question and reuse the cached state exactly. That reuse is checked byte for byte.

How do I build Blink and try a decision quickly?

git clone, then make to build the C runtime, tools and benchmarks with no Python needed. make test runs unit tests plus the no-allocation check, make bench reports latency and memory as JSON, and make install puts blink, libblink, blink.h, a pkg-config file and a man page under /usr/local. Alternative backends build with make ACCELERATE=1 or make W8A8=1.

What are the measured latencies and throughput?

On one core of an Apple M5 Pro, blink-tiny over a 256-byte state takes 406 microseconds for a fresh decision and 54 when the state is reused, at 18,211 decisions per second. W8A8 is 221 and 33 microseconds, and Accelerate is 172 and 32. On x86-64 the AVX2 kernel does the same decision in 760 microseconds fresh and 90 reused.

What shape of decision does the runtime take?

It answers questions of the form given this state and this criterion, which of these options, in a single forward pass. You encode the state once and the option menu once, then scoring is cheap. The output is one probability per option you declared at call time, with no token generation, decoding loop or JSON to parse.

CommunityJev Demos & Experiments

Blink — sqliteai

Embeddable C99 runtime for one-pass typed decisions that returns one probability per option with no token loop, plus a 66 KB WebAssembly build for browsers and Node.

See highlights

Open github.com

Stars: 20
Language: C
License: APACHE-2.0
Published by: Community
Author: sqliteai
Last verified: Oct 2026

sqliteai/blink

Organization repository on GitHub

View on GitHub

An open-source, high-performance System One Model for one-pass typed decisions, with an embeddable C runtime and WebAssembly support.

Forks: 1
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.Homepage

This is one of several Jev listings for the same project. The canonical entry is Blink — ellipsis-dev.

About this project

A one-pass typed-decision model with an embeddable C runtime that also builds to WebAssembly. TypeSafe released Jev, a closed model in a class they call System One Models: fast structured decisions for software rather than conversation, with calibrated confidence and no output tokens to pay for. Blink puts that interface behind a C library you can link into a service, a daemon or a device, or load as WebAssembly in a web page.

Quoted from the project's README. Read the full README.

Tagscommunity jevbest catalog-import c

Works forJev for Classification

Used inJev for Software & DevTools

Highlights

Single forward pass returns one probability per declared option, with no decoding loop or JSON repair step.
Dependency-free C99 runtime with read-only mapped weights and scoring that performs no allocations.
Same sources build to a small WebAssembly module that runs in browsers and Node with no filesystem or server.
State encoding is cached and reused exactly, so follow-up questions on one state cost far less than a fresh decision.

Watch out

Blink judges the form of a decision, not deep reading, and the larger small preset is marked experimental and slower. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

computer-use-cache

Drop-in OpenAI-compatible cache proxy for repeatable agent workflows: exact hits stay local, and with a TypeSafe key Jev judges whether a cached response is semantically reusable for the current request.

Jev Demos & Experiments#community#jevbest#catalog-import

Community

rohanarun/computer-use-cacheJavaScript

open-jev-typed-decision-engine

Open 150M encoder that reproduces the typed noul, choice, and score interface in one forward pass, trainable on a free GPU and runnable locally for nothing.

Jev Demos & Experiments#community#jevbest#catalog-import

Community

intikhab49/open-jev-typed-decision-enginePython

TypeLLM

941

Library that adds schema-guaranteed typed outputs to existing autoregressive LLMs without changing weights, served through SGLang or a hosted API.

Jev Demos & Experiments#community#jevbest#catalog-import

Community

TypeLLM/TypeLLMPython

solar-mini4-jev

Drop-in service that exposes Upstage Solar models through the System One shape, accepting state plus noul, choice, or score questions.

Jev Demos & Experiments#community#jevbest#catalog-import

Community

hunkim/solar-mini4-jevPython

laya-server

Self-hosted Docker service and web UI that packages Laya structured-decision inference behind a compatible System One endpoint.

Jev Demos & Experiments#community#jevbest#catalog-import

Community

1Panel-dev/laya-serverPython

open-medical-jev

Research recipe that fuses two frozen open readers into Jev-style yes/no judgments with a fit-free router, confidence gate, and candidate sets.

Jev Demos & Experiments#community#jevbest#catalog-import

Community

FeiLiuEM/open-medical-jevPython

From the Jev guides

Original write-ups on what this entry is for.

Per-Tick Decisions in Games

Sep 28, 2026

A game tick is a state plus a legal action list. How Clash Royale, Minecraft, and the earlier Mario, Pokémon, StarCraft, and drone loops ask Jev one step at a time.

Jev Demos & Experiments#gaming#agents#control

Read article

Back to all resources

From the Jev community

Posts from builders shipping with Jev right now.

Follow @typesafeai