sam860/
gemma3:270m

90 2 weeks ago

Google's new small model

270m 1b

2 weeks ago

b16d6d39dfbd · 241MB

model
gemma3
·
268M
·
Q4_0
template
{{- $systemPromptAdded := false }} {{- range $i, $_ := .Messages }} {{- $last := eq (len (slice $.Me
license
Gemma Terms of Use Last modified: February 21, 2024 By using, reproducing, modifying, distributing,
params
{ "stop": [ "<end_of_turn>" ], "top_k": 64, "top_p": 0.95 }

Readme

Notes

Uploading Unsloth DynamicQuant2 versions for both the 1B and the new 270m Gemma 3 models. Unsloth’s DQ2 offers better accuracy, especially at higher quants, making these small models more capable.

The default latest, 1b, and 270m tags point to the official Quantization-Aware Trained (QAT) versions, which deliver near-fp16 performance even at smaller quants like Q4_0.

Description

Google’s Gemma 3: lightweight, multimodal models built from the same research as the Gemini family.

Key features include a large context window, multilingual support for over 140 languages, and strong performance for their size. These models can handle both text and image inputs to generate text outputs.

Ideal for on-device tasks, quick summarization, chatbots, and other resource-constrained environments where latency is critical.

References

gemma3-qat on HuggingFace

HuggingFace (Unlsoth-DQ2)

HuggingFace (qat-Unsloth-DQ2)