AnyJev
An open-source library that turns supported language models into Jev-style typed decision scorers without task-specific training.
By Nokia Applied Research · reviewed 2026-09-26
Deep profile checked 2026-09-26 · Nokia Applied Research official repository, PyPI package, changelog, and roadmap
What AnyJev does
AnyJev reads the logits of a compatible language model to produce typed choices and probabilities without generating an answer token by token. It provides a practical way to compare existing open models as decision engines and to keep data local, while inheriting each base model's memory, runtime, license, and calibration limits. The project is early-stage and its Jev-compatible HTTP server and broader backend support remain roadmap items.
Pricing model
Free and open source
Position in the stack
Where AnyJev fits
AnyJev is a local compatibility and evaluation layer that reads candidate logits from an existing language model. The selected base model remains the inference engine and determines memory, speed, language ability, license, and much of the final decision quality.
Typical deployment flow
A practical sequence for evaluating AnyJev before making it part of a permanent stack.
- STEP 1
Choose and pin a base model
Select a supported model and revision whose memory footprint, license, and semantic ability fit the deployment.
- STEP 2
Run the decision benches
Verify Choice, Score, and Noul behavior on the project's examples before adding private labelled cases.
- STEP 3
Wrap it in application policy
Add stable schemas, calibrated thresholds, abstention or review paths, observability, and version pinning.
Best for
- Evaluating an existing open model as a typed decision engine
- Keeping decision inputs and model weights local
- Researching calibration and model-size tradeoffs
Not the right layer for
- A turnkey production API without additional serving work
- Users who do not want to manage a base model runtime
- Assuming one calibration transfers across models or workloads
System fit
Requirements and hardware notes
Application requirements are separate from the memory needed by the selected model and context window.
Check your hardwareMinimum baseline
- Python environment
- A supported base language model
- Enough RAM or VRAM for that complete base model
Recommended baseline
- Pinned model and package revisions
- A labelled calibration and regression set
- An isolated service wrapper for application use
Hardware-specific notes
- AnyJev adds little model weight of its own; the base model determines the main memory footprint.
- Apple Silicon, CPU, and NVIDIA suitability depend on the selected backend and model.
- Large base models can remove the latency and cost advantage expected from a decision-only workflow.
Installation and deployment paths
Choose one path that matches the number of users and the level of operations you can maintain.
Data boundary
What stays local and what may leave
- Local data
- Inputs can remain local when the base model and runtime are both local.
- Network dependencies
- Initial model and package downloads require network access; hosted base-model backends create an external data path.
- Accounts and telemetry
- The library itself is open source, but telemetry and account requirements follow the selected model host and surrounding application.
License checkpoints
- AnyJev is Apache-2.0.
- Every base model, dataset, and runtime keeps its own license and use restrictions.
Operational checkpoints
- Treat v0.x APIs and backend coverage as actively evolving.
- Compare option wording and tokenization before trusting probability differences.
- Build a stable HTTP boundary only after pinning the library, model, tokenizer, and calibration settings.
Capabilities
Capabilities refer to the tool's application layer. Hardware and model support can still depend on a connected inference engine.
Deployment and compatibility
- Deployment
- Desktop / localSelf-hosted
- Platforms
- macOS · Windows · Linux
- Hardware backends
- CPU · Apple Silicon · NVIDIA GPU
- Install methods
- PyPI · Python package · Build from source
Models and integrations
- Model formats
- Transformers · vLLM-compatible models
- Common integrations
- Hugging Face TransformersvLLMPython applicationsDecision benchmarks
Strengths
- Works with existing compatible models without fine-tuning
- Open implementation and reproducible benchmarks
- Local execution follows the selected base model runtime
Limitations
- Early v0.x project with active interface development
- Resource requirements depend on the selected base model
- HTTP compatibility and additional backends are still evolving
Planning checklist
Before you choose AnyJev
Answer these questions with the exact models, hardware, users, and data you expect to operate.
- Is the intended base model supported by the current backend?
- Does direct logit scoring remain stable across labels and option counts?
- What memory and latency does the complete base model require?
- Will the application wait for the planned HTTP compatibility layer or build its own service boundary?
AnyJev FAQ
Is AnyJev a downloadable model?
No. It is a library that turns compatible existing language models into typed decision scorers.
Does AnyJev require fine-tuning?
Its core approach does not require task-specific training, but calibration and evaluation on the target workflow are still necessary.
Can AnyJev replace Laya?
They represent different tradeoffs. AnyJev reuses a general model; Laya uses a small specialist non-autoregressive decision architecture.
Official sources
Use these links to confirm current compatibility and installation requirements.
Related tools
An open-source desktop chat application and local API for running models privately on a personal computer.
A hosted System One decision model that returns typed choices, scores, and yes/no probabilities instead of generated prose.
An open benchmark suite for comparing Jev-class typed decision models across correctness, latency, reliability, and cost.
Execution evidence
Known working recipes using AnyJev
Recipes connect hardware, a model artifact, tools, settings, verification, and a reportable result.
No verified recipe is linked to this record yet.
Compatibility estimates remain available in the planner. A recipe appears here only after its exact stack and verification protocol are documented.