Qabbel
Gemma 4 E4B IT vs Gemma 4 12B It
Qiegħed żewġ mudelli sa erbgħa ma' xulxin fuq l-oqsma li jiddeċiedu deployment: kuntest, prezz, reġjun, liċenzja, lingwi, hosting, u x'hu verifikat.
| Gemma 4 E4B IT | Gemma 4 12B It | |
|---|---|---|
| Fornitur | Google DeepMind | Google DeepMind |
| Parametri | 8B | 12B |
| Tieqa tal-kuntest | 131k | Mhux verifikat: il-fornitur ma ddokumentax din id-dikjarazzjoni. |
| Liċenzja | apache-2.0 | apache-2.0 |
| Modalitajiet | text, vision, audio | text, vision, audio |
| Data tar-rilaxx | 02 Mar 2026 | 23 Mej 2026 |
| Lingwi | Qawwi: en, de, fr, es, it, pt Adekwat: nl, pl, sv, cs, ro, el, uk, ru, ar, hi, zh, ja, ko, tr | Mhux verifikat: il-fornitur ma ddokumentax din id-dikjarazzjoni. |
| Piżijiet miftuħa | iva | iva |
| Kompiti | Spiża baxxa, Viżjoni, Appoġġ multilingwi, Sommarizzazzjoni | Appoġġ multilingwi, Viżjoni, Sommarizzazzjoni |
Fejn żewġ mudelli jvarjaw, din it-tabella turi d-differenza u tieqaf. LLM EU ma tippubblikax badge ta' rebbieħ mingħajr punteġġ dokumentat għal dik il-kompitu, u figura nieqsa tibqa' nieqsa minflok ma ssir żero.
Meta tużah
Gemma 4 E4B IT
Saħħiet
- native audio input in the small-model class
- Apache-2.0 licence
- runs on modest hardware
- text, image and audio in one checkpoint
Limiti
- reasoning scores well below the larger Gemma 4 sizes
- 128K-token window, half of the medium models
- no evaluation has been run by LLM EU
Kompiti
Gemma 4 12B It
Saħħiet
- open weights are published on Hugging Face and can be downloaded without a vendor agreement
- Apache-2.0 licence, so commercial self-hosting is permitted without a separate agreement
- at 12B parameters it fits on a single accelerator
- instruction-tuned checkpoint with a chat template in the repository
- speech recognition runs locally, so audio never leaves the operator premises
Limiti
- catalogue only — LLM EU does not run this model, and nothing on this card implies an LLM EU endpoint
- no evaluation has been run by LLM EU
- the published configuration records no context window, so the card carries none
- at 12B parameters, self-hosting needs a single accelerator with enough memory for the weights
- vision checkpoints are heavier to serve than their parameter count alone suggests, because the image encoder and pre-processing run alongside the language model
Kompiti
Kompiti: Viżjoni, Appoġġ multilingwi, Sommarizzazzjoni