¿Por dónde empiezas?
Parte del hardware que tienes o de los modelos que quieres. Compara capacidad, rendimiento y evidencia antes de desplegar.
Select hardware to see model fit results.
Guides & methodology
Browse the complete catalogs and estimation method when you need them.
Todo el hardware21
Cada clase de equipo tiene su presupuesto, servidor y límite realista.
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.
Abrir guíaMac 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.
Abrir guíaMacBook 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.
Abrir guíaFramework 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.
Abrir guíaRadeon 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.
Abrir guíaNVIDIA 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.
Abrir guíaMac 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.
Abrir guíaMac 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.
Abrir guíaGeForce 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.
Abrir guíaGeForce 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.
Abrir guía8 GB CPU laptop
Small text models with short context. Expect patient, private inference rather than speed.
Abrir guía16 GB CPU desktop
Comfortable with 3B–8B Q4 models; 12B is possible only with reduced context and patience.
Abrir guíaApple silicon · 16 GB
A polished 4B–8B local experience when the OS and apps have enough headroom.
Abrir guíaNVIDIA · 8 GB
The mainstream 4B–8B tier. Some 12B INT4 builds fit tightly with modest context.
Abrir guíaNVIDIA · 12 GB
Strong 8B–14B Q4 territory for a single user.
Abrir guíaNVIDIA · 16 GB
14B models are comfortable; 20B-class low-bit MoE models are a tight upper edge.
Abrir guíaApple silicon · 32 GB
24B–32B Q4 models become practical for one local user.
Abrir guíaNVIDIA · 24 GB
The consumer sweet spot for 27B and 30B-A3B INT4/Q4 models.
Abrir guíaApple silicon · 64 GB
32B models with long context or 70B Q4 with conservative headroom.
Abrir guíaWorkstation · 48 GB
70B Q4 serving begins here, with limited room for cache and concurrency.
Abrir guíaDatacenter · 80 GB
Large MoE models and high-throughput servers, still bounded by quantization and cache.
Abrir guíaTodos los modelos24
Qwen3.8 Flash Next
English- Memoria estimada
- ~108 GB
- Contexto
- 262,144 native; 1M extended
Ling-3.0-flash
English- Memoria estimada
- ~74.4 GB
- Contexto
- Check selected runtime and checkpoint
Nex-N2.5-mini
English- Memoria estimada
- ~21 GB
- Contexto
- 262,144 (config max_position_embeddings)
Qwen3.6 27B
- Memoria estimada
- ~18.5 GB
- Contexto
- 262K native
Qwen3.8 27B
- Memoria estimada
- ~19.5 GB
- Contexto
- 262K native
Qwen3.8 Flash Next REAP-288
- Memoria estimada
- ~40.57 GB
- Contexto
- 7.6K recommended on measured 48 GB Mac
Qwen3.6 35B-A3B
- Memoria estimada
- ~25 GB
- Contexto
- 262K native
DeepSeek V4 Flash
- Memoria estimada
- ~176 GB
- Contexto
- 1M native
Gemma 3 1B
- Memoria estimada
- ~1.4 GB
- Contexto
- 32K
Llama 3.2 3B
- Memoria estimada
- ~2.8 GB
- Contexto
- 128K
Qwen3 4B
- Memoria estimada
- ~3.6 GB
- Contexto
- 32K+
Gemma 3 4B
- Memoria estimada
- ~4.2 GB
- Contexto
- 128K
Qwen3 8B
- Memoria estimada
- ~6.8 GB
- Contexto
- 32K+
Gemma 3 12B
- Memoria estimada
- ~9.4 GB
- Contexto
- 128K
Qwen3 14B
- Memoria estimada
- ~11.2 GB
- Contexto
- 32K+
gpt-oss-20b
- Memoria estimada
- ~16 GB
- Contexto
- 128K
Devstral Small 2 24B
- Memoria estimada
- ~18.5 GB
- Contexto
- 256K
Gemma 3 27B
- Memoria estimada
- ~20.5 GB
- Contexto
- 128K
Qwen3 30B-A3B
- Memoria estimada
- ~22.5 GB
- Contexto
- 32K+
Qwen3 32B
- Memoria estimada
- ~24.5 GB
- Contexto
- 32K+
Llama 3.1 70B
- Memoria estimada
- ~46 GB
- Contexto
- 128K
Llama 4 Scout
- Memoria estimada
- ~78 GB
- Contexto
- 10M advertised
gpt-oss-120b
- Memoria estimada
- ~80 GB
- Contexto
- 128K
Qwen-Image 2.1
English- Memoria estimada
- ~33.1 GB
- Contexto
- Not a token context: one prompt plus up to 10 reference images
Elige la capa de inferencia15
Ollama
One-command local chat and app integration
Abrir guíallama.cpp
GGUF models, CPU/GPU offload, embedded and unusual hardware
Abrir guíaLM Studio
Discovering, downloading and testing models without a terminal
Abrir guíaMLX LM
Native Apple silicon inference, experimentation and fine-tuning
Abrir guíaoMLX
Apple-silicon serving with continuous batching, memory guards, and supported SSD-backed paths
Abrir guíavLLM
Linux GPU servers, concurrency and OpenAI-compatible production APIs
Abrir guíaTensorRT-LLM
Maximum NVIDIA throughput after engine tuning
Abrir guíaSGLang
Multi-request GPU serving with prefix caching and speculative decoding
Abrir guíaLocalAI
One OpenAI-compatible endpoint for text, vision, speech and image models, GPU optional
Abrir guíaLemonade
Ryzen AI, Radeon and Strix Halo PCs, including the XDNA2 NPU, behind one local API
Abrir guíaOpenVINO Model Server
Serving OpenVINO IR models on Intel CPUs, Arc GPUs and NPUs with one API
Abrir guíaKoboldCpp
One-file GGUF chat on Windows, macOS or Linux with no Python install
Abrir guíallamafile
Distributing a whole GGUF model as one runnable executable on Linux, macOS, Windows and BSD
Abrir guíaKTransformers
Very large MoE models with hot experts on the GPU and cold experts in system RAM
Abrir guíaUnsloth
Running GGUF or MLX models from a desktop app that also trains and exports them
Abrir guíaCómo estimamos
Partimos de tamaños y cuantización documentados, reservamos espacio para sistema y ejecución y tratamos la caché como extra. MoE usa parámetros totales.
Separamos parámetros totales, parámetros MoE activos, peso cuantizado, compatibilidad y rendimiento medido.
- 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
Antes de descargar 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.