Directory dei modelli
Ogni scheda qui sotto è una famiglia di modelli AI scaricabile gratuitamente ed eseguibile sul tuo computer. La scheda ti dice chi lo produce, quanto è grande, con quale licenza arriva e in cosa è bravo. Le note delle schede restano in inglese (dati del catalogo). Se sei alle prime armi, il glossario definisce ogni termine, oppure prova il picker per trovare quello adatto alla tua macchina.
Meta AI · United States
Mixture-of-experts architecture. Scout pushes context to 10M tokens, Maverick targets frontier-tier reasoning. Custom license restricts use above 700M MAU and forbids training competitors.
Google DeepMind · United States
Tuned for small-footprint deployment. Strong on consumer GPUs and Apple Silicon. The license is permissive for most commercial use but is not OSI-approved.
Microsoft Research · United States
Small-model series with strong reasoning-per-parameter ratios. MIT licensed, ideal for embedded and on-device scenarios.
Alibaba (Qwen team) · China
The most permissive frontier-class family for smaller sizes. Strong on Asian languages, competitive on English code and reasoning benchmarks.
DeepSeek AI · China
Topped open-source leaderboards on SWE-Bench and GPQA Diamond in early 2026. The MIT-licensed core variant is the most permissive of the top-tier open weights.
Moonshot AI · China
June 2026 update of the K2 line: +21.8% over K2.6 on Kimi Code Bench v2, plus a HighSpeed variant with ~6× faster inference. For the frontier-scale Kimi K3 (2.8T), see the Frontier section below.
Alibaba (Qwen team) · China
Mid-2026 refresh of the Qwen family. Keeps the permissive licensing on smaller sizes that made Qwen the default choice for commercial local deployments, with improved multilingual coverage.
MiniMax · China
Released June 1, 2026: the first open weight to combine frontier coding (vendor-reported 59.0% on SWE-Bench Pro), a 1M-token context and native multimodality in one model. Quantized builds are demanding; check the memory estimates before downloading.
Z.ai (Zhipu) · China
GLM-5.2 (June 2026) is the current #1 open weight on the Artificial Analysis Intelligence Index — but at 744B it belongs in the Frontier section below. The smaller 5.1-era sizes and GLM-4.7-Flash remain the runnable entry points, all MIT licensed.
OpenBMB · China
The edge-first family from OpenBMB (Tsinghua-affiliated). MiniCPM5-1B runs on phones and 8 GB laptops with room to spare, and holds its own against models twice its size. Nearly 400k downloads within two months of release.
01.AI · China
Steady, well-documented series with bilingual strength and reasonable license terms. Less flashy than Qwen or DeepSeek but a solid choice for many production cases.
Mistral AI · France (EU)
Released April 2026. The most credible non-Chinese, non-American option at frontier level. Mistral remains the leading European LLM lab by capability.
EuroLLM Consortium · European Union
Fully open: weights, data pipeline, and training logs. Trained on the MareNostrum 5 supercomputer in Barcelona. Covers all 24 EU official languages plus 11 more. Not at frontier level on English benchmarks, but unique in its transparency.
Allen Institute for AI · United States (non-profit)
The most truly open model in the catalog: weights, data, training code, and intermediate checkpoints are all public. Capability is mid-tier, transparency is unmatched.
Cohere · Canada
Open weights but non-commercial. Strong on retrieval-augmented generation and tool calling. A useful baseline for RAG work even if not deployable as-is in production.
Frontier open weights
Kimi K3, GLM-5.2, Llama 4 Maverick, DeepSeek V4 Pro: cosa serve davvero per eseguirli, e il fratello minore eseguibile di ogni famiglia.
Tendenze · aggiornato ogni settimana
La top 16 live da Hugging Face, ordinata per download, likes e freschezza. Si aggiorna da sola ogni lunedì.
