Jev Local — Argos1111

No. It matches TypeSafe's /v1/systemone shape with local LFM, Sarashina, or ModernBERT backends, and image input is a local extension that does not mean official SDK image compatibility. Check the documented API compatibility range.

Start with LFM text or vision scripts since they need no extra Python packages, then add Sarashina or ModernBERT. Client, eval tools, and API shape stay shared across backends.

The table compares practical use under one API with different training and quantization, not a fair model ranking. ModernBERT trained on JGLUE train scores high there but trails LFM on knowledge tasks, while short Japanese intent tests favor ModernBERT.

The default LFM path keeps images at 512px or smaller with four or fewer questions for speed, accepts PNG/JPEG up to 4 MiB and four images, and reuses no decoder state cache for image questions. ModernBERT rejects images with 422.

Python 3.12 or newer and Bash. The automatic setup covers Linux x86_64 and arm64, macOS on Apple Silicon and Intel, and Windows through WSL2. setup.sh detects the GPU, picks the CUDA, ROCm or Metal llama.cpp build, downloads the model and verifies its SHA256, and falls back to CPU with a stated reason when no GPU is usable; --backend cpu forces that.

LFM adds no Python packages because the binary is fetched automatically, and the vision configuration with the image encoder is about 2.1 GB. Sarashina needs no extra Python dependency for text, with Q4_K_M around 2.07 GB and Q8_0 around 3.57 GB, but its image path needs a manually placed mmproj plus a matching pre-release runtime. ModernBERT needs torch and transformers in its own venv, about 1.3 GB of weights, roughly 2 GB of VRAM, and runs 12 questions in about two seconds on CPU.

Yes, with ./train_modernbert.sh, which pulls the public Japanese datasets automatically (JGLUE train, JCoLA, JCommonsenseMorality and MASSIVE). It takes about 40 minutes on an R9700 in bf16 at roughly 20 GB of VRAM, or about 10 GB with --pair-budget 64, and writes models/modernbert-ja-310m-jev, which run_modernbert.sh then prefers over the published weights.