LOGISTAR — Scheda Oggetto

Dati oggetto

URL

Apri

Titolo / Descrizione

Sito / Azienda

Estratto / Body

&lt;p&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt;Sono emersi diversi modelli linguistici, piccoli ma potenti, a dimostrazione che le dimensioni non sono tutto. I seguenti esempi sono modelli linguistici spaziali (SLM) con un numero di parametri compreso tra 1 e 4 miliardi:&lt;/span&gt;&lt;/span&gt;&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;a href=&quot;https://huggingface-co.translate.goog/meta-llama/Llama-3.2-1B?_x_tr_sl=auto&amp;amp;_x_tr_tl=it&amp;amp;_x_tr_hl=it&quot;&gt;&lt;strong&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt;Llama3.2-1B&lt;/span&gt;&lt;/span&gt;&lt;/strong&gt;&lt;/a&gt;&lt;em&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt; &amp;ndash; Una variante da 1 miliardo di parametri&lt;/span&gt;&lt;/span&gt;&lt;/em&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt; sviluppata da Meta e&lt;/span&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt;ottimizzata per dispositivi edge.&lt;/span&gt;&lt;/span&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://huggingface-co.translate.goog/Qwen/Qwen2.5-1.5B-Instruct?_x_tr_sl=auto&amp;amp;_x_tr_tl=it&amp;amp;_x_tr_hl=it&quot;&gt;&lt;strong&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt;Qwen2.5-1.5B&lt;/span&gt;&lt;/span&gt;&lt;/strong&gt;&lt;/a&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt; &amp;ndash; Un modello di Alibaba progettato per applicazioni multilingue con&lt;/span&gt;&lt;/span&gt;&lt;em&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt; 1,5 miliardi di parametri&lt;/span&gt;&lt;/span&gt;&lt;/em&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt; .&lt;/span&gt;&lt;/span&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://huggingface-co.translate.goog/deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B?_x_tr_sl=auto&amp;amp;_x_tr_tl=it&amp;amp;_x_tr_hl=it&quot;&gt;&lt;strong&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt;DeepSeeek-R1-1.5B&lt;/span&gt;&lt;/span&gt;&lt;/strong&gt;&lt;/a&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt; - Il modello di ragionamento di prima generazione di DeepSeek, derivato da Qwen2.5 e contenente&lt;/span&gt;&lt;/span&gt;&lt;em&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt; 1,5 miliardi di parametri&lt;/span&gt;&lt;/span&gt;&lt;/em&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt; .&lt;/span&gt;&lt;/span&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://huggingface-co.translate.goog/HuggingFaceTB/SmolLM2-1.7B?_x_tr_sl=auto&amp;amp;_x_tr_tl=it&amp;amp;_x_tr_hl=it&quot;&gt;&lt;strong&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt;SmolLM2-1.7B&lt;/span&gt;&lt;/span&gt;&lt;/strong&gt;&lt;/a&gt;&lt;em&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt; &amp;ndash; Da HuggingFaceTB, un modello linguistico &quot;piccolo&quot; ( 1,7 miliardi di parametri&lt;/span&gt;&lt;/span&gt;&lt;/em&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt; ) all&#039;avanguardia&lt;/span&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt;, addestrato su dataset open source specializzati (FineMath, Stack-Edu e SmolTalk).&lt;/span&gt;&lt;/span&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://huggingface-co.translate.goog/microsoft/Phi-3.5-mini-instruct?_x_tr_sl=auto&amp;amp;_x_tr_tl=it&amp;amp;_x_tr_hl=it&quot;&gt;&lt;strong&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt;Phi-3.5-Mini-3.8B&lt;/span&gt;&lt;/span&gt;&lt;/strong&gt;&lt;/a&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt; &amp;ndash; Il modello aperto di Microsoft, piccolo ma potente, con&lt;/span&gt;&lt;/span&gt;&lt;em&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt; 3,8 miliardi di parametri&lt;/span&gt;&lt;/span&gt;&lt;/em&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt; ottimizzato per il ragionamento e la generazione di codice.&lt;/span&gt;&lt;/span&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://huggingface-co.translate.goog/google/gemma-3-4b-it?_x_tr_sl=auto&amp;amp;_x_tr_tl=it&amp;amp;_x_tr_hl=it&quot;&gt;&lt;strong&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt;Gemma3-4B&lt;/span&gt;&lt;/span&gt;&lt;/strong&gt;&lt;/a&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt; - Sviluppato da Google DeepMind, questo modello leggero ma potente, con&lt;/span&gt;&lt;/span&gt;&lt;em&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt; 4 miliardi di parametri,&lt;/span&gt;&lt;/span&gt;&lt;/em&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt; &amp;egrave; multilingue e multimodale.&lt;/span&gt;&lt;/span&gt;&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt;Ecco altri modelli linguistici pi&amp;ugrave; potenti e di piccole dimensioni disponibili: &lt;/span&gt;&lt;/span&gt;&lt;strong&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt;Mistral 7B&lt;/span&gt;&lt;/span&gt;&lt;/strong&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt; , &lt;/span&gt;&lt;/span&gt;&lt;strong&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt;Gemma 9B&lt;/span&gt;&lt;/span&gt;&lt;/strong&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt; e &lt;/span&gt;&lt;/span&gt;&lt;strong&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt;Phi-4 14B&lt;/span&gt;&lt;/span&gt;&lt;/strong&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt;&lt;span style=&quot;vertical-align: inherit;&quot; dir=&quot;auto&quot;&gt; (anche se non sono sicuro che Phi-4 con 14 miliardi di parametri si qualifichi ancora come &quot;piccolo&quot;, ma &amp;egrave; davvero molto capace :)&lt;/span&gt;&lt;/span&gt;&amp;nbsp;&amp;nbsp;&lt;/p&gt;
&lt;p&gt;Altri famosi slm&lt;br /&gt;&lt;br /&gt;&amp;nbsp; &amp;nbsp; Alpaca - Una versione messa a punto di LLaMA, ottimizzata per l&#039;istruzione seguente&lt;br /&gt;&amp;nbsp; &amp;nbsp; Vicuna - Un chatbot open source addestrato da fine-tuning LLaMA&lt;br /&gt;&amp;nbsp; &amp;nbsp; FLAN-T5 Small - Una versione pi&amp;ugrave; piccola del modello FLAN-T5&lt;br /&gt;&amp;nbsp; &amp;nbsp; DistilGPT2 - Una versione distillata di GPT-2&lt;br /&gt;&amp;nbsp; &amp;nbsp; BERT-Mini - Un modello BERT pi&amp;ugrave; piccolo con 4 strati&lt;br /&gt;&lt;br /&gt;&lt;/p&gt;
&lt;div class=&quot;markdown-heading&quot; dir=&quot;auto&quot;&gt;
&lt;h2 class=&quot;heading-element&quot; dir=&quot;auto&quot;&gt;Frameworks and Tools&lt;/h2&gt;
&lt;a id=&quot;user-content-frameworks-and-tools&quot; class=&quot;anchor&quot; href=&quot;https://github.com/slashml/awesome-small-language-models#frameworks-and-tools&quot;&gt;&lt;/a&gt;&lt;/div&gt;
&lt;ul dir=&quot;auto&quot;&gt;
&lt;li&gt;&lt;a href=&quot;https://github.com/huggingface/transformers&quot;&gt;Hugging Face Transformers&lt;/a&gt; - State-of-the-art Natural Language Processing for PyTorch and TensorFlow 2.0&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://github.com/huggingface/peft&quot;&gt;Peft&lt;/a&gt; - Parameter-Efficient Fine-Tuning (PEFT) methods&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://github.com/periflow/periflow&quot;&gt;Periflow&lt;/a&gt; - A framework for deploying large language models&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://github.com/TimDettmers/bitsandbytes&quot;&gt;bitsandbytes&lt;/a&gt; - 8-bit CUDA functions for PyTorch&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://www.tensorflow.org/lite&quot; rel=&quot;nofollow&quot;&gt;TensorFlow Lite&lt;/a&gt; - A set of tools to help developers run TensorFlow models on mobile, embedded, and IoT devices&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://github.com/microsoft/onnxruntime&quot;&gt;ONNX Runtime&lt;/a&gt; - Cross-platform, high performance ML inferencing and training accelerator&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://github.com/always-further/deepfabric&quot;&gt;DeepFabric&lt;/a&gt; - A framework generating high-quality synthetic data for training and evaluating small language models&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;&amp;nbsp;&lt;/p&gt;
&lt;p&gt;&amp;nbsp;&lt;/p&gt;

Note

Voto oggetto (1-10)

Precisione (1-10)

Immagine

Tag (8)

✎ Modifica tag 🤖 Suggerisci AI

          llm:5, ai:5, slm:5, tool:5, opensource:5, embedding:5, llm_internal:5        

ai 5 embedding 5 llm 5 llm_internal 5 opensource 5 slm 5 tool 5 wpg 5

ID: 68124 • Tipo: WPG • Reg: 2026-07-15 • Mod: 29/07/2026 20:16