DeepOpen
Open multilingual System 1 decision engine built on Laya checkpoints with a router that picks the best checkpoint per request.
Overview
DeepOpen describes itself as a fully open-source, non-autoregressive System 1 decision engine built on Convai's Laya, aimed at classification, routing and scoring. It performs multi-dimensional classification across more than 100 languages in a single forward pass and includes a built-in router that matches each request to the most suitable checkpoint, along with presets.
The README is written as a technical whitepaper. It reports latency as low as 33 ms per request and 7.2 ms in batch on NVIDIA T4 GPUs and says the models were trained with the RLCD framework. Reproduction pipelines and checkpoints are provided for the Banking77 and CLINC150 intent benchmarks. According to a list summary, it is essentially a router and presets over Laya checkpoints packaged as its own engine.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which intent or class does this input belong to?(inferred) | — |
| Choice | Which Laya checkpoint should handle this request (router)?(inferred) | — |
How it works
- Receive a classification or scoring request
- Router selects a Laya checkpoint
- Single forward pass returns typed probabilities
- Use results for routing or scoring
Strengths
- Multilingual (100+ languages)
- Low reported latency on T4 GPUs
- Reproduction scripts for intent benchmarks
Caveats
- Marketing-style claims not independently verified
- Largely a wrapper over Laya checkpoints
Getting started
See the banking77 and clinc150 directories in the repository for reproduction scripts.At a glance
- Type
- Model
- Status
- beta
- Language
- Python
- Runtime
- GPU inference engine
- Has code
- yes
- License
- Apache-2.0
Use cases and integrations
How curated lists describe it
A router and presets over Convai's Laya checkpoints, packaged as its own engine.
MrJev/awesome-jev
Frequently asked questions
What is DeepOpen built on?
Convai's Laya checkpoints, with a router and presets.
How fast is it?
The README reports 33 ms per request and 7.2 ms batched on T4 GPUs.
Which benchmarks are reproduced?
Banking77 and CLINC150 intent classification.
Related projects
Japanese System One decision model on ModernBERT-ja (314.6M) returning typed choice, score and bool distributions with no generation.
ProjectbetaPythonOEV: 184M-parameter decision engine that scores options in one pass, reporting 0.7705 on typed-decisions and 22 ms per question.
ProjectresearchPythonFastino's 340M Apache-2.0 decision classifier that scores caller-defined labels in one forward pass for routing, intent and scoring.
ModelproductionPythonMangaba AI's local, calibrated typed-decision model for Portuguese text, trained to optimize the Brier score; proprietary since v5.
ProjectproductionPython- blink★ 22
Open one-pass typed-decision model with a dependency-free C99 runtime that also builds to a 66 KB WebAssembly module.
ServicebetaC - Foq★ 11
Local open alternative to Jev using a 1.58-bit ternary 8B model to answer booleans, choices, scores and Pydantic extraction in ~25 ms.
ProjectbetaPython