jeff
Self-hosted GLiFormer System One server: POST /v1/systemone compatible with typesafe-sdk; local uv run or Modal GPU/CPU deploy.
Overview
jeff (github.com/logan-markewich/jeff, MIT) is Logan Markewich's drop-in System One server compatible with TypeSafe POST /v1/systemone and the official typesafe-sdk when you set TYPESAFE_BASE_URL to your jeff host. It serves knowledgator/gliformer-large-v1 with Choice, Score, and Noul support, temperature-scaled normalized sigmoids, and optional noul isolation modes documented in README. Run locally with uv run jeff or deploy CPU or GPU on Modal. Auth uses JEFF_API_KEYS. Token counts are not comparable to Jev billing per upstream README.
Problem: Teams want System One SDK compatibility without per-call hosted Jev bills, but chat-model JSON hacks break calibration and void comparisons to api.typesafe.ai.
Built for: Python engineers who can run uv locally or deploy Modal and want a self-hosted POST /v1/systemone endpoint backed by the GLiFormer encoder (~400M params).
First indexed on Jev Directory: 2026-09-28
Creator and team
- Name
- Logan Markewich
- Handle
- @logan-markewich
How Jev is used
- Role in the product flow
- Self-hosted System One wire compatibility: same SDK clients, GLiFormer inference instead of hosted Jev
- Primitives
- ChoiceScoreNoul
- State in
- Standard System One JSON bodies from typesafe-sdk or curl, matching hosted Jev request shape per README.
- Decision out
- Typed answers with probabilities or scores; behavior differs from hosted Jev (sigmoid scaling, noul isolation options) as documented upstream.
- Configure TYPESAFE_BASE_URL to jeff host and JEFF_API_KEYS
- uv run jeff locally or deploy Modal template from README
- POST /v1/systemone from existing SDK clients unchanged
- Compare latency, cost, and accuracy using README and bench/RESULTS.md tables
jeff speaks the Jev wire format so existing SDK clients keep working; it does not pretend to be identical hosted Jev. jaredpalmer-kev ships Qwen-trained checkpoints with full fine-tune story; wfzyx-von optimizes non-autoregressive encoder latency; githubnext-localjev proxies chat JSON on Mac; ollaya-dev-ollaya pulls many ONNX decision checkpoints behind one daemon; receptron-laya embeds Convai Laya in Node. jeff is the GLiFormer encoder path with Modal deploy and honest accuracy versus cost tables in README. Hosted Jev remains the accuracy leader in upstream benchmarks; jeff targets operators who accept the gap for roughly 5x lower per-million request cost in README math. Cite logan-markewich/jeff README and bench/RESULTS.md only.
Sourced performance claims
- README benchmarks on 1600 labeled items: sequential p50 about 151 ms on L4 Modal versus hosted Jev about 129 ms.Source: github.com/logan-markewich/jeff README
- README cost table: about $2.6 versus about $15.6 per 1M single-question requests (jeff vs hosted Jev).Source: github.com/logan-markewich/jeff README
- README JevBench v1.2.2 total score 66.9 (rank #9) versus hosted Jev 75.3 (rank #2); AG News 75.5% vs 90.5%; Intelligence 63.9 vs 90.4.Source: github.com/logan-markewich/jeff README and bench/RESULTS.md
- Public GitHub repo logan-markewich/jeff had about 256 stars and 20 forks when this listing was drafted.Source: GitHub API September 2026
Features and stack
Features
- POST /v1/systemone compatible with typesafe-sdk via TYPESAFE_BASE_URL
- knowledgator/gliformer-large-v1 (~400M) GLiFormer backend
- Local uv run jeff or Modal GPU and CPU deploy paths
- JEFF_API_KEYS authentication
- Published latency, cost, and JevBench comparison tables
Stack
- Python
- uv
- GLiFormer weights on Hugging Face
- Modal optional
- typesafe-sdk
Pricing: MIT open source; you pay Modal or your own GPUs. README documents lower per-request cost than hosted Jev with lower benchmark scores.
Links
FAQ
- Is jeff API-identical to hosted Jev?
- Wire compatible for typesafe-sdk clients, but README documents different sigmoid behavior, noul isolation options, and non-comparable token billing.
- How do I point the SDK at jeff?
- Set TYPESAFE_BASE_URL to your jeff server and configure JEFF_API_KEYS per README quick start.
- How does accuracy compare to api.typesafe.ai?
- README and bench/RESULTS.md publish JevBench and task tables where hosted Jev leads; jeff trades accuracy for cost and self-hosting control.
- How is jeff different from Kev?
- Kev (jaredpalmer-kev) fine-tunes Qwen checkpoints with a training and serve story. jeff serves GLiFormer as a drop-in encoder server with Modal deploy.
Related learn guides
Original Jev guidance that pairs with this product pattern.
- System One model
The model family behind Jev: parallel typed questions, one forward pass, probabilities you can threshold.
- Where to run Jev
Compare official TypeSafe, Vercel AI Gateway, OpenRouter, Cloudflare Workers AI, and classifier.dev with a fact table.
Related products
Hand-picked neighbors with rich profiles or overlapping tags.