Wo fängst du an?
Starte mit vorhandener Hardware oder Zielmodellen. Vergleiche Kapazität, Leistung und Belege vor der Installation.
Select hardware to see model fit results.
Guides & methodology
Browse the complete catalogs and estimation method when you need them.
Alle Hardwareklassen21
Jede Geräteklasse hat ein eigenes Speicherbudget, einen Pfad und eine ehrliche Obergrenze.
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.
Leitfaden öffnenMac 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.
Leitfaden öffnenMacBook 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.
Leitfaden öffnenFramework 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.
Leitfaden öffnenRadeon 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.
Leitfaden öffnenNVIDIA 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.
Leitfaden öffnenMac 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.
Leitfaden öffnenMac 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.
Leitfaden öffnenGeForce 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.
Leitfaden öffnenGeForce 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.
Leitfaden öffnen8 GB CPU laptop
Small text models with short context. Expect patient, private inference rather than speed.
Leitfaden öffnen16 GB CPU desktop
Comfortable with 3B–8B Q4 models; 12B is possible only with reduced context and patience.
Leitfaden öffnenApple silicon · 16 GB
A polished 4B–8B local experience when the OS and apps have enough headroom.
Leitfaden öffnenNVIDIA · 8 GB
The mainstream 4B–8B tier. Some 12B INT4 builds fit tightly with modest context.
Leitfaden öffnenNVIDIA · 12 GB
Strong 8B–14B Q4 territory for a single user.
Leitfaden öffnenNVIDIA · 16 GB
14B models are comfortable; 20B-class low-bit MoE models are a tight upper edge.
Leitfaden öffnenApple silicon · 32 GB
24B–32B Q4 models become practical for one local user.
Leitfaden öffnenNVIDIA · 24 GB
The consumer sweet spot for 27B and 30B-A3B INT4/Q4 models.
Leitfaden öffnenApple silicon · 64 GB
32B models with long context or 70B Q4 with conservative headroom.
Leitfaden öffnenWorkstation · 48 GB
70B Q4 serving begins here, with limited room for cache and concurrency.
Leitfaden öffnenDatacenter · 80 GB
Large MoE models and high-throughput servers, still bounded by quantization and cache.
Leitfaden öffnenAlle Modelle24
Qwen3.8 Flash Next
English- Geschätzter Speicher
- ~108 GB
- Kontext
- 262,144 native; 1M extended
Ling-3.0-flash
English- Geschätzter Speicher
- ~74.4 GB
- Kontext
- Check selected runtime and checkpoint
Nex-N2.5-mini
English- Geschätzter Speicher
- ~21 GB
- Kontext
- 262,144 (config max_position_embeddings)
Qwen3.6 27B
- Geschätzter Speicher
- ~18.5 GB
- Kontext
- 262K native
Qwen3.8 27B
- Geschätzter Speicher
- ~19.5 GB
- Kontext
- 262K native
Qwen3.8 Flash Next REAP-288
- Geschätzter Speicher
- ~40.57 GB
- Kontext
- 7.6K recommended on measured 48 GB Mac
Qwen3.6 35B-A3B
- Geschätzter Speicher
- ~25 GB
- Kontext
- 262K native
DeepSeek V4 Flash
- Geschätzter Speicher
- ~176 GB
- Kontext
- 1M native
Gemma 3 1B
- Geschätzter Speicher
- ~1.4 GB
- Kontext
- 32K
Llama 3.2 3B
- Geschätzter Speicher
- ~2.8 GB
- Kontext
- 128K
Qwen3 4B
- Geschätzter Speicher
- ~3.6 GB
- Kontext
- 32K+
Gemma 3 4B
- Geschätzter Speicher
- ~4.2 GB
- Kontext
- 128K
Qwen3 8B
- Geschätzter Speicher
- ~6.8 GB
- Kontext
- 32K+
Gemma 3 12B
- Geschätzter Speicher
- ~9.4 GB
- Kontext
- 128K
Qwen3 14B
- Geschätzter Speicher
- ~11.2 GB
- Kontext
- 32K+
gpt-oss-20b
- Geschätzter Speicher
- ~16 GB
- Kontext
- 128K
Devstral Small 2 24B
- Geschätzter Speicher
- ~18.5 GB
- Kontext
- 256K
Gemma 3 27B
- Geschätzter Speicher
- ~20.5 GB
- Kontext
- 128K
Qwen3 30B-A3B
- Geschätzter Speicher
- ~22.5 GB
- Kontext
- 32K+
Qwen3 32B
- Geschätzter Speicher
- ~24.5 GB
- Kontext
- 32K+
Llama 3.1 70B
- Geschätzter Speicher
- ~46 GB
- Kontext
- 128K
Llama 4 Scout
- Geschätzter Speicher
- ~78 GB
- Kontext
- 10M advertised
gpt-oss-120b
- Geschätzter Speicher
- ~80 GB
- Kontext
- 128K
Qwen-Image 2.1
English- Geschätzter Speicher
- ~33.1 GB
- Kontext
- Not a token context: one prompt plus up to 10 reference images
Die passende Inferenzschicht15
Ollama
One-command local chat and app integration
Leitfaden öffnenllama.cpp
GGUF models, CPU/GPU offload, embedded and unusual hardware
Leitfaden öffnenLM Studio
Discovering, downloading and testing models without a terminal
Leitfaden öffnenMLX LM
Native Apple silicon inference, experimentation and fine-tuning
Leitfaden öffnenoMLX
Apple-silicon serving with continuous batching, memory guards, and supported SSD-backed paths
Leitfaden öffnenvLLM
Linux GPU servers, concurrency and OpenAI-compatible production APIs
Leitfaden öffnenTensorRT-LLM
Maximum NVIDIA throughput after engine tuning
Leitfaden öffnenSGLang
Multi-request GPU serving with prefix caching and speculative decoding
Leitfaden öffnenLocalAI
One OpenAI-compatible endpoint for text, vision, speech and image models, GPU optional
Leitfaden öffnenLemonade
Ryzen AI, Radeon and Strix Halo PCs, including the XDNA2 NPU, behind one local API
Leitfaden öffnenOpenVINO Model Server
Serving OpenVINO IR models on Intel CPUs, Arc GPUs and NPUs with one API
Leitfaden öffnenKoboldCpp
One-file GGUF chat on Windows, macOS or Linux with no Python install
Leitfaden öffnenllamafile
Distributing a whole GGUF model as one runnable executable on Linux, macOS, Windows and BSD
Leitfaden öffnenKTransformers
Very large MoE models with hot experts on the GPU and cold experts in system RAM
Leitfaden öffnenUnsloth
Running GGUF or MLX models from a desktop app that also trains and exports them
Leitfaden öffnenSo schätzen wir
Dokumentierte Größen und Quantisierung bilden die Basis. Wir reservieren Platz für System und Laufzeit; Kontextcache kommt hinzu. MoE nutzt Gesamtparameter.
Gesamtparameter, aktive MoE-Parameter, quantisierte Größe, Laufzeit-Support und Messwerte bleiben getrennt.
- 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
Vor dem 40-GB-Download
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.