Gemma Models

Gemma 4

Google · Gemma

Best for: Developers evaluating Google-backed open-weight models for efficient local apps, hosted prototypes, and multimodal workflows where supported.

Local: Smaller Gemma variants are practical for local testing; larger variants need more VRAM or unified memory.

Gemma 3 4B IT

Google · Gemma

Best for: Efficient local prototypes, app workflows, and Gemma-family comparisons.

Local: Use the exact checkpoint and quantization that matches your hardware and latency target.

Gemma 4 26B (A4B)

Google · Gemma

Best for: High-quality local chat and reasoning on 24 GB cards, and usable split-mode performance below that thanks to sparse activation.

Gemma 4 31B

Google · Gemma

Best for: Maximum Gemma quality for local chat, reasoning, and long-context work on 24 GB cards.

Gemma 3 4B

Google · Gemma 3

Best for: Gemma 3 4B is Google's compact open-weight model, notable for pairing a small 4B parameter count with an unusually large 128K token context window.

Local: 128K context at 4B scale is a standout feature. Q4 fits in 4 GB VRAM. Useful for RAG and summarisation on constrained hardware.

Gemma 4 12B

Google · Gemma 4

Best for: Gemma 4 12B is Google's current flagship open-weight model at the 12B size, pairing a 256K context window with the permissive Apache 2.0 license (a step up from the Gemma Terms of Use that covered the Gemma 3 line).

Local: Google's flagship 12B open-weight model with 256K context. Q4 fits in 8–10 GB VRAM; strong default for MacBooks with 16 GB unified memory where KV-cache growth at long context is manageable. Benchmark against Qwen3 14B on your prompts — Gemma 4 leads on certain reasoning tasks.

Gemma 4 12B (QAT)

Google · Gemma 4

Best for: Gemma 4 12B (QAT) is a quantization-aware trained release of the base Gemma 4 12B model, meaning the model was trained with Q4 quantization already factored in rather than quantized after the fact.

Local: Quantization-aware trained variant of Gemma 4 12B. QAT recovers most of the quality lost to Q4 quantization versus post-training quantization of the base checkpoint, at the cost of a slightly larger Q4_K_M footprint (9.5 GB vs 7.6 GB). QAT is published as a Q4-class artifact only — for Q8 or BF16 inference use the base Gemma 4 12B entry instead.

CodeGemma 7B

Google · Gemma

Best for: Coding assistant experiments and developer workflow prototypes.

Local: Use the exact checkpoint and quantization that matches your hardware and latency target.

Gemma 3 12B

Google · Gemma 3

Best for: Gemma 3 12B is Google's mid-size Gemma 3 checkpoint, built for chat, RAG, and summarization tasks that benefit from its 128K context window without needing a 30B+ model.

Local: 128K context at 12B. Q4 fits in 8–10 GB VRAM with short context; KV cache grows quickly at 128K. Benchmark against Qwen3 14B on your tasks.

PaliGemma 2

Google · Gemma

Best for: Vision-language app prototypes and multimodal evaluation.

Local: Use the exact checkpoint and quantization that matches your hardware and latency target.

Gemma 3 27B

Google · Gemma

Best for: Developers testing capable medium-sized chat models with broad tooling support.

Local: Can be tested locally with quantized builds on higher-end consumer GPUs or unified-memory systems.

Gemma 3 12B IT

Google · Gemma

Best for: Efficient local prototypes, app workflows, and Gemma-family comparisons.

Local: Use the exact checkpoint and quantization that matches your hardware and latency target.

Gemma 3 1B IT

Google · Gemma

Best for: Efficient local prototypes, app workflows, and Gemma-family comparisons.

Local: Use the exact checkpoint and quantization that matches your hardware and latency target.

Gemma 2 27B Instruct

Google · Gemma

Best for: Efficient local prototypes, app workflows, and Gemma-family comparisons.

Local: Use the exact checkpoint and quantization that matches your hardware and latency target.

Gemma 2 9B Instruct

Google · Gemma

Best for: Efficient local prototypes, app workflows, and Gemma-family comparisons.

Local: Use the exact checkpoint and quantization that matches your hardware and latency target.

Gemma 2 2B Instruct

Google · Gemma

Best for: Efficient local prototypes, app workflows, and Gemma-family comparisons.

Local: Use the exact checkpoint and quantization that matches your hardware and latency target.