Da dove inizi?
Parti dall’hardware che possiedi o dai modelli desiderati. Confronta capacità, prestazioni e prove prima del deployment.
Select hardware to see model fit results.
Guides & methodology
Browse the complete catalogs and estimation method when you need them.
Tutto l’hardware21
Ogni classe ha un budget, un runtime e un limite realistico.
Mac Studio M5 Ultra · 96 GB
The entry M5 Ultra configuration combines a 96 GB unified pool with 1.2 TB/s memory bandwidth. Apple announced it in August 2026; availability begins September 22.
Apri guidaMac mini M6 · 24 GB
A compact current-generation Mac with 24 GB unified memory. Keep macOS, runtime, and KV-cache headroom separate from model weights.
Apri guidaMacBook Pro M5 Max · 64 GB
A portable Apple-native tier for larger quantized models. Sustained inference still depends on thermals, runtime support, and context length.
Apri guidaFramework Desktop · Ryzen AI Max+ 395 · 128 GB
A 128 GB Strix Halo desktop. AMD allows a large graphics-memory allocation, but usable capacity and speed depend on BIOS, driver, and runtime support.
Apri guidaRadeon AI PRO R9700 · 32 GB
A 32 GB AMD workstation card for larger single-GPU quantized models. Check the exact operating-system and ROCm or Vulkan runtime path first.
Apri guidaNVIDIA DGX Spark · 128 GB
A compact local AI system with 128 GB coherent memory. It can hold large low-bit models, but decode speed and software support remain model-specific.
Apri guidaMac Studio M3 Ultra · 96 GB
A quiet single-box option for 70B-class Q4 models and Apple-native MLX workflows, with memory shared by macOS and applications.
Apri guidaMac Studio M3 Ultra · 256 GB
Enough capacity for many 100B-plus low-bit models, provided the architecture is supported and optimized by the selected runtime.
Apri guidaGeForce RTX 5090 · 32 GB
A fast single-GPU tier for 27B-class low-bit models, with limited remaining space for very long context or multiple concurrent users.
Apri guidaGeForce RTX 4090 · 24 GB
A 27B Q4 checkpoint can be a tight fit. Leave room for runtime overhead and KV cache instead of filling every gigabyte with weights.
Apri guida8 GB CPU laptop
Small text models with short context. Expect patient, private inference rather than speed.
Apri guida16 GB CPU desktop
Comfortable with 3B–8B Q4 models; 12B is possible only with reduced context and patience.
Apri guidaApple silicon · 16 GB
A polished 4B–8B local experience when the OS and apps have enough headroom.
Apri guidaNVIDIA · 8 GB
The mainstream 4B–8B tier. Some 12B INT4 builds fit tightly with modest context.
Apri guidaNVIDIA · 12 GB
Strong 8B–14B Q4 territory for a single user.
Apri guidaNVIDIA · 16 GB
14B models are comfortable; 20B-class low-bit MoE models are a tight upper edge.
Apri guidaApple silicon · 32 GB
24B–32B Q4 models become practical for one local user.
Apri guidaNVIDIA · 24 GB
The consumer sweet spot for 27B and 30B-A3B INT4/Q4 models.
Apri guidaApple silicon · 64 GB
32B models with long context or 70B Q4 with conservative headroom.
Apri guidaWorkstation · 48 GB
70B Q4 serving begins here, with limited room for cache and concurrency.
Apri guidaDatacenter · 80 GB
Large MoE models and high-throughput servers, still bounded by quantization and cache.
Apri guidaTutti i modelli24
Qwen3.8 Flash Next
English- Memoria stimata
- ~108 GB
- Contesto
- 262,144 native; 1M extended
Ling-3.0-flash
English- Memoria stimata
- ~74.4 GB
- Contesto
- Check selected runtime and checkpoint
Nex-N2.5-mini
English- Memoria stimata
- ~21 GB
- Contesto
- 262,144 (config max_position_embeddings)
Qwen3.6 27B
- Memoria stimata
- ~18.5 GB
- Contesto
- 262K native
Qwen3.8 27B
- Memoria stimata
- ~19.5 GB
- Contesto
- 262K native
Qwen3.8 Flash Next REAP-288
- Memoria stimata
- ~40.57 GB
- Contesto
- 7.6K recommended on measured 48 GB Mac
Qwen3.6 35B-A3B
- Memoria stimata
- ~25 GB
- Contesto
- 262K native
DeepSeek V4 Flash
- Memoria stimata
- ~176 GB
- Contesto
- 1M native
Gemma 3 1B
- Memoria stimata
- ~1.4 GB
- Contesto
- 32K
Llama 3.2 3B
- Memoria stimata
- ~2.8 GB
- Contesto
- 128K
Qwen3 4B
- Memoria stimata
- ~3.6 GB
- Contesto
- 32K+
Gemma 3 4B
- Memoria stimata
- ~4.2 GB
- Contesto
- 128K
Qwen3 8B
- Memoria stimata
- ~6.8 GB
- Contesto
- 32K+
Gemma 3 12B
- Memoria stimata
- ~9.4 GB
- Contesto
- 128K
Qwen3 14B
- Memoria stimata
- ~11.2 GB
- Contesto
- 32K+
gpt-oss-20b
- Memoria stimata
- ~16 GB
- Contesto
- 128K
Devstral Small 2 24B
- Memoria stimata
- ~18.5 GB
- Contesto
- 256K
Gemma 3 27B
- Memoria stimata
- ~20.5 GB
- Contesto
- 128K
Qwen3 30B-A3B
- Memoria stimata
- ~22.5 GB
- Contesto
- 32K+
Qwen3 32B
- Memoria stimata
- ~24.5 GB
- Contesto
- 32K+
Llama 3.1 70B
- Memoria stimata
- ~46 GB
- Contesto
- 128K
Llama 4 Scout
- Memoria stimata
- ~78 GB
- Contesto
- 10M advertised
gpt-oss-120b
- Memoria stimata
- ~80 GB
- Contesto
- 128K
Qwen-Image 2.1
English- Memoria stimata
- ~33.1 GB
- Contesto
- Not a token context: one prompt plus up to 10 reference images
Scegli il livello di inferenza15
Ollama
One-command local chat and app integration
Apri guidallama.cpp
GGUF models, CPU/GPU offload, embedded and unusual hardware
Apri guidaLM Studio
Discovering, downloading and testing models without a terminal
Apri guidaMLX LM
Native Apple silicon inference, experimentation and fine-tuning
Apri guidaoMLX
Apple-silicon serving with continuous batching, memory guards, and supported SSD-backed paths
Apri guidavLLM
Linux GPU servers, concurrency and OpenAI-compatible production APIs
Apri guidaTensorRT-LLM
Maximum NVIDIA throughput after engine tuning
Apri guidaSGLang
Multi-request GPU serving with prefix caching and speculative decoding
Apri guidaLocalAI
One OpenAI-compatible endpoint for text, vision, speech and image models, GPU optional
Apri guidaLemonade
Ryzen AI, Radeon and Strix Halo PCs, including the XDNA2 NPU, behind one local API
Apri guidaOpenVINO Model Server
Serving OpenVINO IR models on Intel CPUs, Arc GPUs and NPUs with one API
Apri guidaKoboldCpp
One-file GGUF chat on Windows, macOS or Linux with no Python install
Apri guidallamafile
Distributing a whole GGUF model as one runnable executable on Linux, macOS, Windows and BSD
Apri guidaKTransformers
Very large MoE models with hot experts on the GPU and cold experts in system RAM
Apri guidaUnsloth
Running GGUF or MLX models from a desktop app that also trains and exports them
Apri guidaCome stimiamo
Partiamo da dimensioni e quantizzazione documentate, riserviamo spazio al sistema e trattiamo la cache come extra. MoE usa i parametri totali.
Separiamo parametri totali, parametri MoE attivi, peso quantizzato, supporto e prestazioni misurate.
- Memory fit is an estimate, not a speed measurement.
- Size MoE by total parameters; shared and distributed memory need additional checks.
- Vendor claims, community measurements, runtime conditions and unknowns are disclosed separately.
Catalog data updated: 2026-10-03
Prima di scaricare 40 GB
Does a 24 GB GPU run a 30B model?
Often at Q4/INT4 with a conservative context. Qwen3 30B-A3B and Gemma 3 27B are representative fits, but cache and runtime overhead still matter.
Are active MoE parameters the memory requirement?
No. Active parameters affect compute per token; total parameters still need to be stored in memory or offloaded.
Which runtime should a beginner choose?
Ollama for a terminal-first setup or LM Studio for a visual desktop. llama.cpp is the portable fallback; vLLM is for higher-throughput GPU serving.
How do you report speed?
Exact device-and-model measurements are labeled measured. Relative manufacturer figures are labeled vendor claims. Every other combination stays unknown instead of receiving an invented tokens-per-second estimate.