Skip to content
Metatext Decide — typed decisions for your agents, with a probability on every answerGet early access
Metatext
Sign Up

Intel-Phi-Jev

Local server (xks) implementing Jev's /v1/systemone contract on an open-weight LLM, accelerated with Xeon Phi cards.

Projectresearch
View on GitHub ↗

Overview

Intel Phi Jev provides xks, a local implementation of TypeSafe's Jev contract. Clients send state and typed Noul, Choice or Score questions to POST /v1/systemone and get probability distributions and confidence, never generated text, so the official SDKs and the companion Mechanical Jev client work by pointing TYPESAFE_BASE_URL at it. The model is an open-weight LLM on the machine (Qwen3.8-35B-A3B by default), read from its next-token distribution, with arithmetic offloaded to two Xeon Phi 3120 cards via a sibling AVX-512 project.

It is a hobbyist and research system for people who want Jev-compatible decisions on their own hardware. The state is rendered once as a session and each question becomes a suffix whose next token is an option label; user text is tokenized with special tokens off so it cannot break out of the document. Because it cannot replicate TypeSafe's training, it calibrates readings after the fact. Make targets cover install, doctor checks, serving, queries, experiments and a full check with tests.

What the decision model decides

TypeQuestionAnswers
OtherAny Noul, Choice or Score question sent in Jev's wire format—

How it works

  1. Run xks doctor and make serve to start the local server
  2. Client posts state and questions to /v1/systemone
  3. State is prefilled once; each question suffix reads option-label probabilities
  4. Readings are calibrated and returned in Jev's format

Strengths

  • Drop-in local endpoint for official Jev SDKs
  • No data leaves the machine
  • Prompt construction prevents caller text from opening chat turns

Caveats

  • Requires unusual hardware (Xeon Phi cards) for its intended setup
  • Uses post-hoc calibration rather than Jev's training
  • Hobby-scale project; accuracy relative to Jev not stated

Getting started

`make install`, `xks doctor`, `make serve`, then `curl -s localhost:8090/v1/systemone -d @examples/query.json`.

At a glance

Type
Project
Status
research
Language
Rust
Runtime
local HTTP server on llama.cpp
Has code
yes
Has tests
yes

Use cases and integrations

How curated lists describe it

Frequently asked questions

Is this TypeSafe's Jev?

No. It is a local open-weight LLM served behind Jev's API contract.

Will official Jev SDKs work with it?

Yes, by setting TYPESAFE_BASE_URL to the xks server.

Do I need Xeon Phi cards?

They are the project's target accelerator; plain x86-64 llama.cpp builds are also supported.

Related projects

  • ifReflex.cpp

    llama.cpp-based C++ server that turns any LLM into a System One decision model, reading answers from next-token logits.

    ProjectbetaC++
  • JEV-MLX

    MLX library that turns an existing local model into a decision engine by reading candidate logits, with no fine-tuning, on Apple Silicon.

    ProjectprototypePython
  • jev-bush

    Educational C11 library and CLI that turns DiffusionGemma 26B into a CPU-first OpenJev-style decision engine reading candidate logits.

    ProjectresearchC
  • jevify (Mintzs)

    Experimental engine that makes an existing LLM do Jev-like classification and scoring by reusing context and building JSON.

    ProjectresearchPython
  • open-bonsai-jev

    Reads typed decisions from one forward pass of PrismML's 1.75-bit Bonsai 27B under llama.cpp, using the OpenJev/SemIf logit method.

    ProjectresearchPython
  • rwkv-jev

    Jev-class decision server on an RWKV-7 1B checkpoint that reads typed answers from a constrained softmax behind /v1/systemone.

    ProjectresearchPython