Kev 1.0
An open family of Jev-like direct-decision models spanning 0.8B to 27B parameters, with public code and model artifacts.
Kev 1.0 is a community project whose maintainer describes the models as open System One and Jev-like alternatives. It is not an official TypeSafe release or a verified drop-in replacement for Jev.
- Provider
- Jared Palmer / Kev
- Reviewed
- October 8, 2026
- Project status
- Community
- Verification
- Verified
- Categories
- automation, development, research
- Compatibility
- Python, PyTorch, CUDA, Apple MLX, Hugging Face
Inputs
- ✓Text context
- ✓Direct-decision questions and candidate options
Outputs
- ✓Typed decisions
- ✓Candidate probabilities
- ✓Confidence values
Interface
- ✓structured-output
Decision tasks
- ✓Classify
- ✓Choose
- ✓Score
- ✓Rank
- ✓Route
Verified model facts
Provider facts and directory notes are separated. Time-sensitive claims include attribution and should be checked again before production use.
Community project. Variant specifications and licensing statements are maintainer-reported. Kev is not affiliated with TypeSafe and is not presented here as an official or drop-in Jev replacement.
- Current verified version
- Kev 1.0
- Model family
- Kev direct-decision models
- Release date
- 2026-10-01
- Access methods
- local-package, open-weights
- Hosted API
- No
- Local deployment
- Yes
- Downloadable weights
- Yes
- License
- Apache-2.0 code; maintainer states Apache-2.0 for released weights
- Supported languages
- Language behavior depends on the selected Qwen base model and Kev training data
Commercial-use note
The repository code is Apache-2.0 and the maintainer states the released weights use Apache-2.0. Each Qwen base model and dependency retains separate provenance and terms that must be reviewed before commercial use.
Context or input limits
The 0.8B, 4B, and 9B v2 variants report 8,192-token validation and 65,536-token serving limits; the 27B v2 variant reports 65,536 for both. Smaller training states reached 7,552 tokens, while 27B training states reached 32,768.
Pricing
No per-request software price is listed for local use. Compute, storage, and hosting costs depend on the selected variant and deployment environment.
Latency and rate limits
The reviewed release does not provide one hardware-normalized latency figure that applies across all four variants. Benchmark the chosen artifact on production hardware.
Strengths
- ✓Four published sizes for local evaluation and deployment.
- ✓Public code with CUDA and Apple MLX paths.
- ✓Separates validated context from the larger serving limit.
Limitations
- ✓Maintainer descriptions and measurements have not been independently reproduced by Type Safe.
- ✓A serving context limit does not prove equal quality across the full window.
- ✓The 27B bf16 artifact is about 51 GB and requires substantial storage and memory.
- ✓Base-model licenses and provenance remain separate from the Kev repository license.
Safety notes
- ✓Validate every variant on representative tasks before selecting thresholds.
- ✓Keep a human-review path for low-confidence or consequential decisions.
Model variants
Validated context is the source-reported tested range. Serving context is the runtime acceptance limit; the two values are not interchangeable.
| Variant | Parameters | Base model | Artifact | Validated context | Serving context |
|---|---|---|---|---|---|
| Kev-0.8B | 0.8B | Qwen3.5-0.8B-Base | LoRA adapter and decision head | 8,192 tokens | 65,536 tokens |
| Kev-4B | 4B | Qwen3.5-4B-Base | LoRA adapter and decision head | 8,192 tokens | 65,536 tokens |
| Kev-9B v2 | 9B | Qwen3.5-9B-Base | LoRA adapter and decision head | 8,192 tokens | 65,536 tokens |
| Kev-27B v2 | 27B | Qwen3.8-27B | Post-trained full bf16 model and decision head | 65,536 tokens | 65,536 tokens |
Implementation guidance
Keep the declared inputs and outputs visible in code review. Validate every model response before it reaches a database, API, or customer-facing action.