36 models · ranked by HuggingFace downloads
Gemma 4-E4B-IT is Google DeepMind's edge-optimized 4-billion-parameter any-to-any multimodal model from the Gemma 4 family, designed for deployment on mobile and edge devices rather than servers. The 'any-to-any' pipeline_tag indicates multimodal input and output capability beyond standard image-text-to-text. Apache 2.0 licensed.
Gemma 4 E2B is Google's efficient 2B-parameter multimodal model, instruction-tuned for both image-text and text-only prompts. It targets edge and on-device deployment where a sub-3B footprint is necessary.
gemma-4-12B-it is Google's Gemma 4 multimodal (text + image) instruction-tuned model. It accepts both text and image inputs and produces text, making it suitable for document analysis, visual Q&A, and structured data extraction. Released under Apache-2.0, it targets users who need a capable VLM without access restrictions.
gemma-4-12B-it-qat-w4a16-ct is a quantization-aware trained (QAT) weights for W4A16 deployment version of Google's Gemma 4 multimodal (text + image) instruction-tuned model. 12B parameters are reduced to lower-precision weights for deployment on memory-constrained hardware or Apple Silicon, with quality degradation typically small for general chat tasks. The base model is Apache-2.0 licensed.
Nemotron-3-Nano-Omni-30B-A3B-Reasoning-NVFP4 processes and generates across multiple modalities, enabling cross-modal reasoning in a single model call.
A 4-bit MLX quantization of Google's Gemma 4 E4B instruct model (an efficient 4B-equivalent MoE variant) for Apple Silicon. Targets developers who want Gemma 4 running locally on MacBook-class hardware.
gemma-4-12B-it-qat-GGUF is Unsloth's GGUF repack of Google's Gemma 4 12B instruction-tuned model, which was originally quantization-aware trained (QAT) at q4_0. GGUF packaging enables CPU and hybrid CPU/GPU inference via llama.cpp-compatible runtimes without requiring a full PyTorch stack. The Apache 2.0 license and Unsloth's optimization focus make this a practical option for local inference on consumer hardware.
An 8-bit MLX quantization of Google's Gemma 4 E4B instruct model for Apple Silicon. Higher quality than the 4-bit variant at the cost of roughly double the memory, targeting M2/M3 Pro or Max class machines.
gemma-4-E4B-it-MLX-5bit is a MLX 5-bit quantized weights optimized for Apple Silicon inference version of Google's Gemma 4 MoE-based multimodal (text + image) instruction-tuned model. parameters are reduced to lower-precision weights for deployment on memory-constrained hardware or Apple Silicon, with quality degradation typically small for general chat tasks. The base model is Apache-2.0 licensed.
gemma-4-E4B-it-MLX-6bit is a MLX 6-bit quantized weights optimized for Apple Silicon inference version of Google's Gemma 4 MoE-based multimodal (text + image) instruction-tuned model. parameters are reduced to lower-precision weights for deployment on memory-constrained hardware or Apple Silicon, with quality degradation typically small for general chat tasks. The base model is Apache-2.0 licensed.
As a minicpm-based open-weight model, MiniCPM-o-4_5 focuses on multimodal any-to-any generation. The Apache 2.0 license keeps MiniCPM-o-4_5 unrestricted for commercial reuse. Check the MiniCPM-o-4_5 model card for benchmarks and intended use before adopting it.
Qwen3-Omni-30B-A3B-Instruct handles multiple input and output modalities including text, images, and audio within a single unified architecture.
NVIDIA's Nemotron-H Nano Omni 30B-A3B is a multimodal reasoning-focused any-to-any model quantized to FP8-dynamic. The 'Nano Omni' designation signals a compact MoE footprint (3B active) designed for reasoning tasks spanning text, image, and code modalities.
Qwen2.5-Omni-3B handles multiple input and output modalities including text, images, and audio within a single unified architecture.
This is the official Google release of Gemma 4 12B instruction-tuned in GGUF format, quantized to q4_0 using Quantization-Aware Training. Unlike community repacks, this comes directly from Google, providing clearer provenance for production pipelines that require verified model sources.
Google's QAT W4A16 (4-bit weights, 16-bit activations) quantized Gemma 4 E4B instruction-tuned model, packaged with compressed-tensors. QAT W4A16 is a quantization scheme that preserves activation precision while compressing weights, balancing accuracy and throughput.
gemma-4-31B-it-assistant is Google's 31-billion-parameter instruction-tuned Gemma 4 model configured for assistant-style interactions. Listed under the any-to-any pipeline tag, it is designed to handle flexible input-output modality combinations within the Transformers ecosystem.
A GGUF conversion of Google's Gemma-4-E4B-IT with QAT (Quantization-Aware Training) applied, then further quantized for local deployment. QAT bakes quantization error into the training objective, making the resulting model more robust to low-bit inference than post-training quantization. The any-to-any pipeline tag indicates multimodal capability.
Gemma 4-E4B is Google's 4-billion effective-parameter instruction-tuned model from the Gemma 4 family, quantization-aware trained (QAT) to Q4_0 precision and packaged as GGUF. QAT differs from post-training quantization by incorporating quantization error into the training objective, generally preserving more quality at 4-bit than naive PTQ. Gemma 4 supports multi-modal any-to-any inputs.
A dynamically FP8-quantized variant of Google's Gemma 4-12B instruction-tuned model, produced by Red Hat AI. Dynamic FP8 applies per-tensor scaling at runtime rather than during calibration, making it simple to produce but slightly less efficient than static-calibrated FP8. Red Hat AI packages models for enterprise OpenShift AI deployments.
Unsloth's GGUF distribution of Google's Gemma 4 E2B instruction-tuned model with QAT (Quantization-Aware Training). E2B is the 2-billion effective-parameter variant — the smallest of the Gemma 4 family. QAT bakes quantization error into training, preserving quality better than post-training quantization at this aggressively small scale.
gemma-4-E2B-it-qat-q4_0-gguf is a quantization-aware-training (QAT) GGUF export of Google's Gemma-4 E2B instruct model at q4_0 precision. QAT typically recovers accuracy lost in post-training quantization, making this preferable to a naive PTQ conversion at the same bit-width.
Google's official QAT w4a16 (4-bit weights, 16-bit activations) compressed-tensors version of Gemma-4-E2B-it. QAT during training generally preserves more accuracy than post-training quantization at the same bit width, making this a high-quality INT4 option for multimodal Gemma-4.
gemma-4-E4B is a multimodal model accepting diverse input types and producing outputs across text, vision, and audio modalities.
OneThinker-SFT is a Qwen3-8B model fine-tuned by OneThink with supervised fine-tuning (SFT) on a vision-language task mixture, using the Qwen3-VL architecture for any-to-any multimodal output. Apache-2.0 licensed.
As a qwen3-based large model, Qwen3-Omni-30B-A3B-Thinking focuses on multimodal any-to-any generation. Weighing in near 30000M parameters, Qwen3-Omni-30B-A3B-Thinking trades some ceiling for cheaper, faster inference. Qwen3-Omni-30B-A3B-Thinking lists a non-standard license, so confirm permissions before deployment. Read Qwen3-Omni-30B-A3B-Thinking's card for hardware requirements and licensing fine print before deploying.
Google's quantization-aware trained (QAT) Gemma 4 12B instruction-tuned model using Q4_0 scheme, released in its unquantized (BF16) form for post-hoc quantization. QAT-trained models are fine-tuned to tolerate 4-bit weight quantization better than post-training-only quantization.
Nemotron-3 Nano Omni is NVIDIA's multimodal reasoning model — 30B total parameters with 3B active per token — that extends the Nemotron-H architecture to support any-to-any input and output modalities including audio, image, and text. The Reasoning variant includes a thinking mode for extended chain-of-thought. It runs in BF16 full precision, targeting multi-GPU H100/H200 deployments.
Built for multimodal any-to-any generation, gemma-4-12B is a gemma-based model with publicly available weights. At about 12000M parameters, gemma-4-12B sits in the large tier, which sets its memory and latency budget. gemma-4-12B is Apache 2.0-licensed, clearing it for closed-source and paid products. gemma-4-12B ships without a hosted SLA, so budget for self-managed deployment and monitoring.
MiniCPM-o 2.6 is an omnimodal 8B model from OpenBMB supporting speech, image, and text inputs with real-time audio output. It targets on-device multimodal scenarios, particularly mobile and edge deployments, with end-to-end speech conversation capability.
An uncensored 'Heretic' fine-tune of Google's Gemma 4-E4B instruction-tuned model by llmfan46. The Heretic alignment style targets broad instruction compliance by removing safety constraints, applied here to Gemma 4's efficient 4B-effective-parameter multimodal architecture. Any-to-any input capability is retained from the base.
An AWQ INT4 quantization of Google's Gemma-4-E4B-IT, a multimodal any-to-any model. AWQ applies activation-aware weight quantization to minimize accuracy degradation at INT4 bit depth. This variant targets GPU deployments where the full model memory footprint is too large.
Community AWQ INT4 quantization of Google's Gemma-4-12B multimodal instruction-tuned model, supporting image and text input. Uses compressed-tensors format, enabling Gemma-4 12B to run on mid-range GPUs at reduced VRAM cost.
Qwen2.5-Omni-7B is a multimodal model accepting diverse input types and producing outputs across text, vision, and audio modalities.
gemma-4-E4B-it-assistant is an openly licensed multimodal any-to-any generation model in the gemma family. At about 4000M parameters, gemma-4-E4B-it-assistant sits in the mid-sized tier, which sets its memory and latency budget. gemma-4-E4B-it-assistant is Apache 2.0-licensed, clearing it for closed-source and paid products. Treat gemma-4-E4B-it-assistant's published metrics as a starting point and validate against your workload.
Gemma-4-E2B is Google's 2B edge model from the Gemma-4 family, designed for on-device deployment with multimodal any-to-any capability. The 'E' prefix indicates edge-optimized — smaller memory footprint and lower latency are prioritized over raw capability. Supports image and text input/output in a single model.