User repository on GitHub
A self-hosted drop-in replacement for TypeSafe's jev, powered by GliFormer.
- Forks
- 24
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
Jev directory
CommunityJev Practices & Patterns
Jeff — logan-markewich
A self-hosted implementation of TypeSafe's Jev System One API powered by the 400M-parameter GLiFormer encoder: it serves choice, score, and noul and drops into the official typesafe-sdk via TYPESAFE_BASE_URL, but trails Jev on reasoning-heavy tasks.
- Stars
- 303
- Language
- Python
- License
- MIT
- Reading time
- 3 min
- Published by
- Community
- Author
- logan-markewich
- Last verified
- Sep 2026
This is one of several Jev listings for the same project. The canonical entry is Jeff — firelex.
About this project
A self-hosted implementation of TypeSafe's jev System One API, powered by GLiFormer (400M parameters). Use the official typesafe-sdk by pointing TYPESAFEBASEURL at jeff. Supports choice (pick an option), score (rate on ordered levels), and noul (probability of yes). Cheaper to self-host, but less accurate than jev on reasoning-heavy tasks. See benchmarks. Quickstart · Deploy · Configuration · API · Development. Requires uv and Python 3.12. Run from the repository root:
Quoted from the project's README. Read the full README.
Used inJev for Software & DevTools
Highlights
- Serves choice, score, and noul on a 400M-parameter GLiFormer encoder; the official typesafe-sdk works by pointing TYPESAFE_BASE_URL at it.
- Measured HTTP throughput caps at about 50 requests/s per Modal container, and an A10G outperforms an L4 for direct backend calls and long requests.
- Its benchmark write-up is close on binary sentiment and tied on emotion, but behind Jev on irony, reading comprehension, and the judge and hard tiers.
- Request limits default to 64 questions, 64 labels, and 20,000 state characters, with per-key rate limiting and batching at 16 items and 5 ms waits.
- Quickstart needs Python 3.12 and uv; the server picks CUDA, then MPS, then CPU, and the CPU fallback is reported slower and more expensive than Jev.
Quickstart
bash
uv sync --extra dev
uv run hf download knowledgator/gliformer-large-v1 --local-dir models/gliformer-large-v1
JEFF_API_KEYS=devkey uv run jeffWatch out
MIT-licensed. Needs Python 3.12, uv, and a local GLiFormer checkpoint downloaded from Hugging Face; hosted cost figures are estimates that depend on deployment throughput.
Questions about this entry
Short answers drawn from the linked project or page.
Yes, that is the point. You point TYPESAFE_BASE_URL at jeff and the official typesafe-sdk talks to it, answering choice, score and noul against a self-hosted GLiFormer encoder. The README is explicit that the wire format works while the model behavior differs.
Not exactly. Probabilities are normalized sigmoids temperature-scaled at 3.2, so a score matches the weighted average of displayed probabilities only at JEFF_TEMPERATURE=1. Nouls get separate encoder passes while choice and score share one pass and can affect each other; JEFF_ISOLATE=all buys independence at extra cost.
No. Measured on 1,600 labeled items across eight datasets, AG News topic accuracy is 75.5 percent against 90.5, and on JevBench Intelligence jeff scores 63.9 across easy, standard, judge and hard versus 90.4. Its JevBench rank comes from cost; on intelligence alone it is 14th of 18.
Per the README, no. An 8-core Modal CPU deployment with an int8 ONNX encoder measured slower and more expensive than Jev, and ONNX runs only the encoder while the RNN and classification head stay in PyTorch. The L4 HTTP path measured about 50 requests/s per container at roughly $2.6 per 1M single-question requests against $15.6 for Jev.
With uv and Python 3.12, run uv sync --extra dev, download gliformer-large-v1 into models/, then JEFF_API_KEYS=devkey uv run jeff; it serves on http://localhost:8000 and selects CUDA, then MPS, then CPU.
By default 64 questions, 64 labels, and 20,000 state characters, with a queue of 256 before HTTP 529; exceeding a limit returns 422, and the batcher defaults to 16 requests and a 5 ms wait.
Noul questions get separate encoder passes, while choice and score questions share a pass and can affect each other; JEFF_ISOLATE=all restores independence at extra cost.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
X: Standard One SH (StandardThinking)
LinkedIn: What multimodal decision models actually change, and what they do not.
LinkedInJevDirectory.org
Hacker News: Decision Index 0.2.1 scores open reproductions of Jev on a frozen suite.
jackbravo4 pointsDecision Index 0.2.1 – 70 open reproductions of TypeSafe Jev's Decision ModelRead the thread on Hacker NewsHacker News: Pointer to a DiffusionGemma plus vLLM tutorial for a Jev-shaped classifier.
1 pointHow to build a Jev-style classifier with DiffusionGemma and vLLMRead the thread on Hacker NewsMore Jev resources like this
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.
Jev Practices & Patterns#community#python#open-models
Community
jaredpalmer/kevPython
jaredpalmer/kevPythonAn 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.
Jev Practices & Patterns#community#python#open-models
Community
Mapika/deciderPython
Mapika/deciderPythonIndependent cross-model benchmark for Jev-class decision models, running 534 frozen cases per complete entrant with scoring code and a four-axis score of accuracy, calibration, latency and cost.
Jev Practices & Patterns#community#python#benchmarks
Community
fstandhartinger/jevbenchPython
fstandhartinger/jevbenchPythonA from-scratch one-pass option scorer that maps a context and a changing list of text options to one probability per option, adding a hashed n-gram encoder, rival-aware attention, a gated head, and temperature scaling; JSONL data is drop-in compatible with jevlike.
Jev Practices & Patterns#community#python#calibration
Community
olanotolu/jevbetterPython
olanotolu/jevbetterPythonA benchmark that runs Jev on all 6,257 text traces of Who&When Pro to attribute agent failures, scoring it with the official pinned scorer against the paper's GPT-5.4, Claude Sonnet 4.6, GLM-5, and Qwen3.5-122B baselines.
Jev Practices & Patterns#community#python#benchmarks
A pre-registered, independently rescored audit of Jev's calibration and type safety: calibrated on CLINC150 (ECE 0.0204), systematically overconfident on Banking77 (ECE 0.0936), and zero type errors across 8,576 responses.
Jev Practices & Patterns#community#python#calibration
Community
jourdanlabs/assay-001Python
jourdanlabs/assay-001PythonFrom the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.