This is a standalone language-only checkpoint based on florianvoss/Gemma-4-E4B-it-TextOnly, prepared as a Hugging Face causal-language-model checkpoint for LLiMa compilation and deployment on Sima.ai hardware. Vision and audio components are not included.

This checkpoint remains subject to the source model's license, intended-use guidance, and limitations.

The retained language quantization used the first 512 HuggingFaceH4/ultrachat_200k train_sft conversations, rendered with the Gemma chat template, truncated to 1,024 tokens, and processed in deterministic dataset order.

The included standalone quantize.py reproduces the prequantize model directly from the Gemma4 TextOnly source. It targets all 344 Linear modules, including lm_head, with GPTQ INT4/G256, static act-order, block size 128, and dampening fraction 0.01. It contains the Gemma4 shared-KV replay fix required during sequential GPTQ calibration and has no vision/audio loading or targeting code.

Transformers compressed loading, finite-scale validation, and greedy text-generation smoke tests passed. Slurm job 4952 produced a coherent factual answer with finite logits. The final conversion validation checked 344 finite quantization scales and generated a coherent joke.

Create a fresh direct-GPTQ checkpoint from the unquantized flattened TextOnly source:

See versions.txt for the exact Python, CUDA, Torch, Transformers, llm-compressor, and compressed-tensors versions.

This is the pre-LLiMa quantized Hugging Face artifact. Compile it for the target Sima.ai platform with the appropriate LLiMa release and keep compiler output separate. No upload is authorized as part of the current local batch.

This artifact contains no vision or audio capability. Quantization quality can vary by language, domain, prompt format, context length, safety requirements, and deployment runtime; validate the intended workload before use.