--- title: "YandexGPT (language model) (IT)" source: "https://systems-analysis.info/int/YandexGPT_(language_model)_(IT)" wiki: "systems-analysis.info/int" article: "YandexGPT_(language_model)_(IT)" language: "it" categories: - "Category:Italian" - "Category:Large language models" - "Category:LLM families" - "Category:Machine learning" revision_id: 8520 wiki_created_at: 2026-09-07T01:18:35Z wiki_modified_at: 2026-09-07T01:18:35Z downloaded_at: 2026-09-07T23:25:50Z --- # YandexGPT (language model) (IT) **YandexGPT (Yet another GPT)** — famiglia di grandi modelli linguistici sviluppati da Yandex e presentati per la prima volta nel maggio 2023.^{[\[1\]](https://systems-analysis.info/int/YandexGPT_(language_model)_(IT)#cite_note-rbc-may17-1)} Le reti YandexGPT sono utilizzate nell'assistente integrato Alisa, nella Ricerca e in altri servizi, e sono disponibili tramite l'API pubblica della piattaforma Yandex Cloud.^{[\[2\]](https://systems-analysis.info/int/YandexGPT_(language_model)_(IT)#cite_note-cloud-api-2)} **YaLM‑100B** (2022) — modello di ricerca precedente con 100 miliardi di parametri e codice aperto; ha rappresentato una «prova di concetto», ma YandexGPT è stato sviluppato separatamente per uso commerciale.^{[\[3\]](https://systems-analysis.info/int/YandexGPT_(language_model)_(IT)#cite_note-github-yalm-3)} ## Storia delle versioni | Data | Versione | Caratteristiche principali | |-------------|---------------------------|---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------| | Giu 2022 | YaLM‑100B | 100 B parametri, 1,7 TB di dati; Apache 2.0.^{[\[3\]](https://systems-analysis.info/int/YandexGPT_(language_model)_(IT)#cite_note-github-yalm-3)} | | 17 mag 2023 | YandexGPT 1.0 | Integrazione in «Alisa».^{[\[1\]](https://systems-analysis.info/int/YandexGPT_(language_model)_(IT)#cite_note-rbc-may17-1)} | | 7 set 2023 | YandexGPT 2 | +67 % di qualità secondo i test interni.^{[\[4\]](https://systems-analysis.info/int/YandexGPT_(language_model)_(IT)#cite_note-rbc-sep8-4)} | | 28 mar 2024 | YandexGPT 3 Pro / Lite | Nuova gamma API aziendale.^{[\[5\]](https://systems-analysis.info/int/YandexGPT_(language_model)_(IT)#cite_note-rbc-mar28-5)} | | 24 ott 2024 | YandexGPT 4 Pro / Lite | Contesto di 32 000 token; ragionamento implicito (chain‑of‑thought).^{[\[6\]](https://systems-analysis.info/int/YandexGPT_(language_model)_(IT)#cite_note-habr-gpt4-6)} | | 25 feb 2025 | YandexGPT 5 Pro | Parità con GPT‑4o nel 64 % dei compiti.^{[\[7\]](https://systems-analysis.info/int/YandexGPT_(language_model)_(IT)#cite_note-adindex-5pro-7)} | | 31 mar 2025 | YandexGPT 5 Lite Instruct | Modello da 8 miliardi di parametri in accesso aperto; formato Llama.^{[\[8\]](https://systems-analysis.info/int/YandexGPT_(language_model)_(IT)#cite_note-hf-5lite-8)} | Versioni principali ## Architettura e addestramento - Architettura di base: transformer ottimizzato per la lingua russa. - YandexGPT 5 Lite: compatibile con Llama; pre-training ≈ 15 trilioni di token, fine‑tuning successivo ≈ 320 miliardi.^{[\[8\]](https://systems-analysis.info/int/YandexGPT_(language_model)_(IT)#cite_note-hf-5lite-8)} ### Contesto e limiti - Limite architetturale del contesto — 32 000 token (versioni 4/5).^{[\[6\]](https://systems-analysis.info/int/YandexGPT_(language_model)_(IT)#cite_note-habr-gpt4-6)} - L'API pubblica limita una singola richiesta (prompt + completion) a 7 400 token.^{[\[9\]](https://systems-analysis.info/int/YandexGPT_(language_model)_(IT)#cite_note-langchain-7400-9)} - La dimensione massima della \*\*risposta\*\* è di 2 000 token secondo la sezione «Quotas and limits».^{[\[10\]](https://systems-analysis.info/int/YandexGPT_(language_model)_(IT)#cite_note-cloud-limits-10)} ## Modelli attuali (giugno 2025) | Modello | Parametri | Contesto | Licenza | Note | |------------------|-----------|----------|-------------------------|--------------------------------------------------------------------------------------------------------------------------------------------------| | YandexGPT 5 Pro | n/d | 32 000 | proprietaria | Accesso tramite API e «Alisa Pro».^{[\[7\]](https://systems-analysis.info/int/YandexGPT_(language_model)_(IT)#cite_note-adindex-5pro-7)} | | YandexGPT 5 Lite | 8 mld | 32 000 | Yandex GPT‑Lite License | Aperto; compatibile con Llama.^{[\[8\]](https://systems-analysis.info/int/YandexGPT_(language_model)_(IT)#cite_note-hf-5lite-8)} | | YaLM‑100B | 100 mld | 2 048 | Apache 2.0 | Progetto originale.^{[\[3\]](https://systems-analysis.info/int/YandexGPT_(language_model)_(IT)#cite_note-github-yalm-3)} | ## Benchmark - Test interni: 5 Pro ha raggiunto la parità con GPT‑4o nel 64 % dei compiti; superamento rispetto a 4 Pro — 67 %.^{[\[7\]](https://systems-analysis.info/int/YandexGPT_(language_model)_(IT)#cite_note-adindex-5pro-7)} - ru‑LLM Arena: YandexGPT mantiene la leadership per punteggio ELO tra i modelli in lingua russa.^{[\[11\]](https://systems-analysis.info/int/YandexGPT_(language_model)_(IT)#cite_note-llmarena-11)} ## Fine‑tuning Per 5 Lite è ufficialmente supportato il metodo LoRA; un esempio di esecuzione è pubblicato nella model card.^{[\[8\]](https://systems-analysis.info/int/YandexGPT_(language_model)_(IT)#cite_note-hf-5lite-8)} ## Modalità API - Sincrona — risposte rapide (Lite). - Asincrona — compiti ad alta intensità di risorse (Pro).^{[\[2\]](https://systems-analysis.info/int/YandexGPT_(language_model)_(IT)#cite_note-cloud-api-2)} ## Multimodalità La famiglia YandexGPT rimane testuale; i servizi multimodali («Neuro», «YandexArt», «Yandex Vision») sono sviluppati separatamente.^{[\[6\]](https://systems-analysis.info/int/YandexGPT_(language_model)_(IT)#cite_note-habr-gpt4-6)} ## Collegamenti esterni - YandexGPT in Yandex Cloud — pagina del servizio - Pesi di YandexGPT‑5 Lite su Hugging Face - Repository di YaLM‑100B su GitHub ## Bibliografia - Matkin, N. et al. (2024). *Comparative Analysis of Encoder-Based NER and Large Language Models for Skill Extraction from Russian Job Vacancies*. arXiv:2407.19816. - Tsanda, A.; Bruches, E. (2024). *Russian-Language Multimodal Dataset for Automatic Summarization of Scientific Papers*. arXiv:2405.07886. - Goloburda, M. et al. (2025). *Qorǵau: Evaluating LLM Safety in Kazakh-Russian Bilingual Contexts*. arXiv:2502.13640. - Togmanov, M. et al. (2025). *KazMMLU: Evaluating Language Models on Kazakh, Russian, and Regional Knowledge of Kazakhstan*. arXiv:2502.12829. - Noels, S. et al. (2025). *What Large Language Models Do Not Talk About: An Empirical Study of Moderation and Censorship Practices*. arXiv:2504.03803. ## Note 1. ↑ ^{[1.0](https://systems-analysis.info/int/YandexGPT_(language_model)_(IT)#cite_ref-rbc-may17_1-0)} ^{[1.1](https://systems-analysis.info/int/YandexGPT_(language_model)_(IT)#cite_ref-rbc-may17_1-1)} «“Яндекс” добавил в “Алису” аналог ChatGPT». *РБК*. [1] 2. ↑ ^{[2.0](https://systems-analysis.info/int/YandexGPT_(language_model)_(IT)#cite_ref-cloud-api_2-0)} ^{[2.1](https://systems-analysis.info/int/YandexGPT_(language_model)_(IT)#cite_ref-cloud-api_2-1)} «Getting started with YandexGPT (Quickstart)». *Yandex Cloud Docs*. [2] 3. ↑ ^{[3.0](https://systems-analysis.info/int/YandexGPT_(language_model)_(IT)#cite_ref-github-yalm_3-0)} ^{[3.1](https://systems-analysis.info/int/YandexGPT_(language_model)_(IT)#cite_ref-github-yalm_3-1)} ^{[3.2](https://systems-analysis.info/int/YandexGPT_(language_model)_(IT)#cite_ref-github-yalm_3-2)} «yandex/YaLM‑100B: Pretrained language model with 100B». *GitHub*. [3] 4. [↑](https://systems-analysis.info/int/YandexGPT_(language_model)_(IT)#cite_ref-rbc-sep8_4-0) «Как “Яндекс” решил зарабатывать на своём аналоге ChatGPT». *РБК*. [4] 5. [↑](https://systems-analysis.info/int/YandexGPT_(language_model)_(IT)#cite_ref-rbc-mar28_5-0) «“Яндекс” представил третье поколение нейросетей YandexGPT». *РБК*. [5] 6. ↑ ^{[6.0](https://systems-analysis.info/int/YandexGPT_(language_model)_(IT)#cite_ref-habr-gpt4_6-0)} ^{[6.1](https://systems-analysis.info/int/YandexGPT_(language_model)_(IT)#cite_ref-habr-gpt4_6-1)} ^{[6.2](https://systems-analysis.info/int/YandexGPT_(language_model)_(IT)#cite_ref-habr-gpt4_6-2)} «Более мощное семейство моделей YandexGPT 4». *Habr*. [6] 7. ↑ ^{[7.0](https://systems-analysis.info/int/YandexGPT_(language_model)_(IT)#cite_ref-adindex-5pro_7-0)} ^{[7.1](https://systems-analysis.info/int/YandexGPT_(language_model)_(IT)#cite_ref-adindex-5pro_7-1)} ^{[7.2](https://systems-analysis.info/int/YandexGPT_(language_model)_(IT)#cite_ref-adindex-5pro_7-2)} «“Яндекс” внедрил YandexGPT 5 Pro в чат с “Алисой Про”». *AdIndex*. [7] 8. ↑ ^{[8.0](https://systems-analysis.info/int/YandexGPT_(language_model)_(IT)#cite_ref-hf-5lite_8-0)} ^{[8.1](https://systems-analysis.info/int/YandexGPT_(language_model)_(IT)#cite_ref-hf-5lite_8-1)} ^{[8.2](https://systems-analysis.info/int/YandexGPT_(language_model)_(IT)#cite_ref-hf-5lite_8-2)} ^{[8.3](https://systems-analysis.info/int/YandexGPT_(language_model)_(IT)#cite_ref-hf-5lite_8-3)} «yandex/YandexGPT‑5‑Lite‑8B‑pretrain». *Hugging Face*. [8] 9. [↑](https://systems-analysis.info/int/YandexGPT_(language_model)_(IT)#cite_ref-langchain-7400_9-0) «ChatYandexGPT API Reference (max_tokens = 7400)». *LangChain Docs*. [9] 10. [↑](https://systems-analysis.info/int/YandexGPT_(language_model)_(IT)#cite_ref-cloud-limits_10-0) «Yandex Cloud service quotas and limits → Foundation Models». *Yandex Cloud Docs*. [10] 11. [↑](https://systems-analysis.info/int/YandexGPT_(language_model)_(IT)#cite_ref-llmarena_11-0) «llmarena/llmarena — российская краудсорсинговая платформа оценки LLM». *GitHub*. [11]