Which JevStyle model should I start with?
Start with 0.8B v3 for a smaller text-only baseline. Choose 2B v4 with PyTorch when the input includes images, or evaluate it as a larger text model. Both score predefined options rather than generate free-form text. Neither choice guarantees accuracy on a new task.
For ticket labels, workflow routes, yes/no checks, and ordered scores, first define the decisions you need. Then measure both releases on the same held-out requests. A larger model's published score alone does not determine its quality, memory use, or speed in your application.
0.8B v3 versus 2B v4
| Property | 0.8B v3 | 2B v4 |
|---|---|---|
| Package release selector | --release 0.8b; package default | --release 2b |
| Model size | 0.8B class; 752M text-model parameters | 2.2B including the vision tower |
| Inputs | Text and JSON | Text and JSON; images with PyTorch only |
| Whole-request token limit | 25,600 | 32,768, including image tokens |
| Output | Probabilities for yes/no, predefined choices, or ordered score levels; no generated explanation | |
| Safetensors weights | 1.50 GB | 4.43 GB |
| MLX bf16 / 8-bit weights | 1.50 / 0.80 GB | 3.76 / 2.00 GB; text only |
| GGUF F16 / Q8_0 / Q4_K_M weights | 1.52 / 0.81 / 0.53 GB | 3.78 / 2.01 / 1.27 GB; text only |
| Published JevBench public accuracy | 64.1% (148 / 231) | 74.9% (173 / 231) |
| License | Apache-2.0; consult the model repository's LICENSE and NOTICE | |
Sizes are published weight-file totals in decimal GB, not peak RAM/VRAM, default downloads, or storage for every backend. Runtime state, inputs, framework overhead, and image processing need additional memory. The project recommends Q8_0 or F16 for long 2B text inputs; a smaller quantized file is not automatically equivalent in quality.
Sources: 0.8B v3 card, 2B v4 card, and the package's build table. Read the benchmark protocols and uncertainty before treating the accuracy rows as a ranking.
Choose the backend for your hardware
| Environment | Starting point | What to check |
|---|---|---|
| Apple silicon, text | jev-style[mlx]; MLX is selected automatically | Use 0.8B for a smaller baseline or compare 2B on your data. Image input is not available in the MLX builds. |
| CUDA GPU, text or images | jev-style[torch] | Only 2B v4 reads images. Measure with the precision and concurrency you plan to deploy. |
| CPU | jev-style[torch] with --device cpu | Measure latency on your machine; the published A100 results do not predict CPU speed. |
| GGUF runtime | Follow the package's backend setup | Use the matching decision scorer. Generic text-generation tools do not produce these decision probabilities. |
Run a local baseline
Use Python 3.10 or newer in a virtual environment. For Apple silicon text input:
python -m pip install "jev-style[mlx]"
jev-style serve --release 0.8bFor the 2B v4 PyTorch build, including image input:
python -m pip install "jev-style[torch]"
jev-style serve --release 2b --backend torchThe first run downloads the model. Once the server is ready, open the local Playground. To compare releases, keep your input set and decision definitions fixed, use separate ports, and follow the evaluation guide.
The hosted API is a separate service with its own available model, input cap, and service status. Do not apply these local limits to it; consult API documentation.
Model selection questions
Can I use JevStyle offline?
Local inference can run without sending decision inputs to a hosted model after the required package, weights, and runtime files have been downloaded. The browser demo and hosted API are separate online services.
Is a 0.53 GB file the total memory requirement?
No. It is the 0.8B Q4_K_M GGUF weight file. Runtime memory is larger and depends on the backend and input. The default package setup does not necessarily choose that file.
Can the MLX or GGUF version of 2B v4 read images?
No. Those distributions contain the language model only. Choose the 2B v4 PyTorch distribution and matching runtime for image input.
Should I always choose the larger model?
No. Compare task accuracy, calibration, warm latency, memory, and review coverage on your own examples. Keep a person or existing LLM available for ambiguous cases and runtime failures.