Lev opens typed decisions on Qwen3.5

// 1h agoOPENSOURCE RELEASE

Lev opens typed decisions on Qwen3.5

Lev is an Apache-2.0 LoRA adapter for Qwen3.5-4B that returns calibrated probabilities for yes/no, choice, and score questions in one forward pass. It supports local GPU deployment and the Jev-compatible `/v1/systemone` protocol. [Interfaze](https://interfaze.ai/blog/jev-now-open-source-lev), [Hugging Face](https://huggingface.co/interfaze-ai/lev)

// ANALYSIS

Lev’s real breakthrough is deployment control: it makes Jev-style typed decisions available as an open, self-hosted primitive, even if its benchmark performance still trails Jev.

–Batched inference answers multiple questions with zero generated tokens, avoiding JSON parsing, retries, and out-of-schema labels.
–The Apache-2.0 adapter runs on Qwen3.5-4B and lets existing Jev clients target a local `/v1/systemone` server.
–Reported S1Bench performance is 68.9% macro versus Jev’s 76.1%, with notable weaknesses on minimal-edit and summary-faithfulness tasks.
–Routing, moderation, triage, and intent detection are the strongest use cases; developers should validate calibration on their own data before automating consequential decisions.
–Real-time use still requires a GPU, and the current release is English-only. [ModelSystem.One](https://modelsystem.one/models/lev/)

// TAGS

levllmsmall-llmopen-weightsstructured-outputinferencelocal-firstopen-source

DISCOVERED

1h ago

2026-09-27

PUBLISHED

1h ago

2026-09-27

RELEVANCE

8/ 10

AUTHOR

DIY Smart Code

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

TUTORIAL1h ago

Harvard's MLSysBook Teaches AI Systems Engineering

Harvard CS249r’s free Machine Learning Systems curriculum covers the engineering behind practical AI, from hardware, memory, and data movement to training, inference, deployment, and responsible systems. It includes open textbooks, labs, TinyTorch, hardware kits, and instructor materials.

BENCHMARK1h ago

JevBench maps typed-decision model tradeoffs.

JevBench v1 evaluates Jev-class decision models across accuracy, cost, latency, calibration, reliability, and openness. Its 19 September run tested nine systems on 242 typed decisions each, with GPT-5.6 Luna and Jev 1.13.0 leading accuracy. [Benchmark Heaven](https://benchmarkheaven.com/jev-models/v1)

OPEN SOURCE2h ago

Sentry Skills packages production engineering workflows

Sentry Skills is an open-source collection of reusable Agent Skills covering code review, branching, commits, security, prompt optimization, documentation, and PR workflows. It installs through Claude Code or the skills.sh CLI and works across compatible coding agents. [GitHub](https://github.com/getsentry/skills)