Which version should I choose?

For a first try, use the web experience or free HTTP API. No model download is needed. For offline use, start with 0.8B; evaluate your own task before moving to 2B or 4B.

Goal	Recommended path
Online trial	Web experience or HTTP API: the shortest trial path
Python / JavaScript applications	HTTP SDK with an existing service; for offline inference, ONNX SDK with 0.8B FP32 and cached local assets
C++ local HTTP service	GGUF Q8_0 with the Valen custom llama-server Linux x86-64 CPU prebuilt package; see GGUF
GPU deployment	Native weights with the repository SGLang/vLLM plugins, or Python ONNX cuda-fp16; follow each backend's measured scope
Fine-tuning	Complete native Valen weights with the repository LoRA/full SFT recipes; see training

0.8B/2B/4B denote model size. Native weights, ONNX and GGUF are file formats. FP32/F32, Q8_0 and ONNX q8-mlp denote format-specific precision/quantization variants, not interchangeable names. Transformers, ONNX Runtime, custom llama.cpp/Ollama and SGLang/vLLM are execution backends. SDK Valen.local loads ONNX, not GGUF; access a GGUF server through HTTP.

All three GGUF sizes publish qualified F32, Q8_0 and matching vision modules. Q4/Q5 failed the quality gates and are not recommended. Custom llama.cpp/Ollama real-model validation covers Linux CPU, not GPU, Windows or other platforms. File sizes describe downloads/disk usage, not peak RAM. No unmeasured throughput or latency is promised.