Par où commencez-vous ?
Partez du matériel possédé ou des modèles visés. Comparez capacité, performances et preuves avant le déploiement.
Select hardware to see model fit results.
Guides & methodology
Browse the complete catalogs and estimation method when you need them.
Tout le matériel21
Chaque classe a son budget mémoire, son moteur et sa limite réaliste.
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.
Ouvrir le guideMac 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.
Ouvrir le guideMacBook 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.
Ouvrir le guideFramework 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.
Ouvrir le guideRadeon 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.
Ouvrir le guideNVIDIA 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.
Ouvrir le guideMac 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.
Ouvrir le guideMac 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.
Ouvrir le guideGeForce 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.
Ouvrir le guideGeForce 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.
Ouvrir le guide8 GB CPU laptop
Small text models with short context. Expect patient, private inference rather than speed.
Ouvrir le guide16 GB CPU desktop
Comfortable with 3B–8B Q4 models; 12B is possible only with reduced context and patience.
Ouvrir le guideApple silicon · 16 GB
A polished 4B–8B local experience when the OS and apps have enough headroom.
Ouvrir le guideNVIDIA · 8 GB
The mainstream 4B–8B tier. Some 12B INT4 builds fit tightly with modest context.
Ouvrir le guideNVIDIA · 12 GB
Strong 8B–14B Q4 territory for a single user.
Ouvrir le guideNVIDIA · 16 GB
14B models are comfortable; 20B-class low-bit MoE models are a tight upper edge.
Ouvrir le guideApple silicon · 32 GB
24B–32B Q4 models become practical for one local user.
Ouvrir le guideNVIDIA · 24 GB
The consumer sweet spot for 27B and 30B-A3B INT4/Q4 models.
Ouvrir le guideApple silicon · 64 GB
32B models with long context or 70B Q4 with conservative headroom.
Ouvrir le guideWorkstation · 48 GB
70B Q4 serving begins here, with limited room for cache and concurrency.
Ouvrir le guideDatacenter · 80 GB
Large MoE models and high-throughput servers, still bounded by quantization and cache.
Ouvrir le guideTous les modèles24
Qwen3.8 Flash Next
English- Mémoire estimée
- ~108 GB
- Contexte
- 262,144 native; 1M extended
Ling-3.0-flash
English- Mémoire estimée
- ~74.4 GB
- Contexte
- Check selected runtime and checkpoint
Nex-N2.5-mini
English- Mémoire estimée
- ~21 GB
- Contexte
- 262,144 (config max_position_embeddings)
Qwen3.6 27B
- Mémoire estimée
- ~18.5 GB
- Contexte
- 262K native
Qwen3.8 27B
- Mémoire estimée
- ~19.5 GB
- Contexte
- 262K native
Qwen3.8 Flash Next REAP-288
- Mémoire estimée
- ~40.57 GB
- Contexte
- 7.6K recommended on measured 48 GB Mac
Qwen3.6 35B-A3B
- Mémoire estimée
- ~25 GB
- Contexte
- 262K native
DeepSeek V4 Flash
- Mémoire estimée
- ~176 GB
- Contexte
- 1M native
Gemma 3 1B
- Mémoire estimée
- ~1.4 GB
- Contexte
- 32K
Llama 3.2 3B
- Mémoire estimée
- ~2.8 GB
- Contexte
- 128K
Qwen3 4B
- Mémoire estimée
- ~3.6 GB
- Contexte
- 32K+
Gemma 3 4B
- Mémoire estimée
- ~4.2 GB
- Contexte
- 128K
Qwen3 8B
- Mémoire estimée
- ~6.8 GB
- Contexte
- 32K+
Gemma 3 12B
- Mémoire estimée
- ~9.4 GB
- Contexte
- 128K
Qwen3 14B
- Mémoire estimée
- ~11.2 GB
- Contexte
- 32K+
gpt-oss-20b
- Mémoire estimée
- ~16 GB
- Contexte
- 128K
Devstral Small 2 24B
- Mémoire estimée
- ~18.5 GB
- Contexte
- 256K
Gemma 3 27B
- Mémoire estimée
- ~20.5 GB
- Contexte
- 128K
Qwen3 30B-A3B
- Mémoire estimée
- ~22.5 GB
- Contexte
- 32K+
Qwen3 32B
- Mémoire estimée
- ~24.5 GB
- Contexte
- 32K+
Llama 3.1 70B
- Mémoire estimée
- ~46 GB
- Contexte
- 128K
Llama 4 Scout
- Mémoire estimée
- ~78 GB
- Contexte
- 10M advertised
gpt-oss-120b
- Mémoire estimée
- ~80 GB
- Contexte
- 128K
Qwen-Image 2.1
English- Mémoire estimée
- ~33.1 GB
- Contexte
- Not a token context: one prompt plus up to 10 reference images
Choisir la couche d’inférence15
Ollama
One-command local chat and app integration
Ouvrir le guidellama.cpp
GGUF models, CPU/GPU offload, embedded and unusual hardware
Ouvrir le guideLM Studio
Discovering, downloading and testing models without a terminal
Ouvrir le guideMLX LM
Native Apple silicon inference, experimentation and fine-tuning
Ouvrir le guideoMLX
Apple-silicon serving with continuous batching, memory guards, and supported SSD-backed paths
Ouvrir le guidevLLM
Linux GPU servers, concurrency and OpenAI-compatible production APIs
Ouvrir le guideTensorRT-LLM
Maximum NVIDIA throughput after engine tuning
Ouvrir le guideSGLang
Multi-request GPU serving with prefix caching and speculative decoding
Ouvrir le guideLocalAI
One OpenAI-compatible endpoint for text, vision, speech and image models, GPU optional
Ouvrir le guideLemonade
Ryzen AI, Radeon and Strix Halo PCs, including the XDNA2 NPU, behind one local API
Ouvrir le guideOpenVINO Model Server
Serving OpenVINO IR models on Intel CPUs, Arc GPUs and NPUs with one API
Ouvrir le guideKoboldCpp
One-file GGUF chat on Windows, macOS or Linux with no Python install
Ouvrir le guidellamafile
Distributing a whole GGUF model as one runnable executable on Linux, macOS, Windows and BSD
Ouvrir le guideKTransformers
Very large MoE models with hot experts on the GPU and cold experts in system RAM
Ouvrir le guideUnsloth
Running GGUF or MLX models from a desktop app that also trains and exports them
Ouvrir le guideNotre méthode
Nous partons des tailles et quantifications documentées, réservons de la mémoire au système et traitons le cache comme un coût en plus. MoE utilise le total.
Nous séparons paramètres totaux, paramètres MoE actifs, poids quantifiés, support et performances mesurées.
- 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
Avant de télécharger 40 Go
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.