Find the model your hardware can run.
Choose your hardware, check model fit, then plan deployment. Find a viable match before downloading or upgrading.
Select hardware to see model fit results.
Guides & methodology
Browse the complete catalogs and estimation method when you need them.
All hardware guides21
Each device class has a different memory budget, runtime path, and honest upper edge.
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.
Open 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.
Open 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.
Open 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.
Open 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.
Open 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.
Open 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.
Open 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.
Open 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.
Open 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.
Open guide8 GB CPU laptop
Small text models with short context. Expect patient, private inference rather than speed.
Open guide16 GB CPU desktop
Comfortable with 3B–8B Q4 models; 12B is possible only with reduced context and patience.
Open guideApple silicon · 16 GB
A polished 4B–8B local experience when the OS and apps have enough headroom.
Open guideNVIDIA · 8 GB
The mainstream 4B–8B tier. Some 12B INT4 builds fit tightly with modest context.
Open guideNVIDIA · 12 GB
Strong 8B–14B Q4 territory for a single user.
Open guideNVIDIA · 16 GB
14B models are comfortable; 20B-class low-bit MoE models are a tight upper edge.
Open guideApple silicon · 32 GB
24B–32B Q4 models become practical for one local user.
Open guideNVIDIA · 24 GB
The consumer sweet spot for 27B and 30B-A3B INT4/Q4 models.
Open guideApple silicon · 64 GB
32B models with long context or 70B Q4 with conservative headroom.
Open guideWorkstation · 48 GB
70B Q4 serving begins here, with limited room for cache and concurrency.
Open guideDatacenter · 80 GB
Large MoE models and high-throughput servers, still bounded by quantization and cache.
Open guideAll model guides24
Qwen3.8 Flash Next
- Estimated memory
- ~108 GB
- Context
- 262,144 native; 1M extended
Ling-3.0-flash
- Estimated memory
- ~74.4 GB
- Context
- Check selected runtime and checkpoint
Nex-N2.5-mini
- Estimated memory
- ~21 GB
- Context
- 262,144 (config max_position_embeddings)
Qwen3.6 27B
- Estimated memory
- ~18.5 GB
- Context
- 262K native
Qwen3.8 27B
- Estimated memory
- ~19.5 GB
- Context
- 262K native
Qwen3.8 Flash Next REAP-288
- Estimated memory
- ~40.57 GB
- Context
- 7.6K recommended on measured 48 GB Mac
Qwen3.6 35B-A3B
- Estimated memory
- ~25 GB
- Context
- 262K native
DeepSeek V4 Flash
- Estimated memory
- ~176 GB
- Context
- 1M native
Gemma 3 1B
- Estimated memory
- ~1.4 GB
- Context
- 32K
Llama 3.2 3B
- Estimated memory
- ~2.8 GB
- Context
- 128K
Qwen3 4B
- Estimated memory
- ~3.6 GB
- Context
- 32K+
Gemma 3 4B
- Estimated memory
- ~4.2 GB
- Context
- 128K
Qwen3 8B
- Estimated memory
- ~6.8 GB
- Context
- 32K+
Gemma 3 12B
- Estimated memory
- ~9.4 GB
- Context
- 128K
Qwen3 14B
- Estimated memory
- ~11.2 GB
- Context
- 32K+
gpt-oss-20b
- Estimated memory
- ~16 GB
- Context
- 128K
Devstral Small 2 24B
- Estimated memory
- ~18.5 GB
- Context
- 256K
Gemma 3 27B
- Estimated memory
- ~20.5 GB
- Context
- 128K
Qwen3 30B-A3B
- Estimated memory
- ~22.5 GB
- Context
- 32K+
Qwen3 32B
- Estimated memory
- ~24.5 GB
- Context
- 32K+
Llama 3.1 70B
- Estimated memory
- ~46 GB
- Context
- 128K
Llama 4 Scout
- Estimated memory
- ~78 GB
- Context
- 10M advertised
gpt-oss-120b
- Estimated memory
- ~80 GB
- Context
- 128K
Qwen-Image 2.1
- Estimated memory
- ~33.1 GB
- Context
- Not a token context: one prompt plus up to 10 reference images
Pick the serving layer for the job15
Ollama
One-command local chat and app integration
Open guidellama.cpp
GGUF models, CPU/GPU offload, embedded and unusual hardware
Open guideLM Studio
Discovering, downloading and testing models without a terminal
Open guideMLX LM
Native Apple silicon inference, experimentation and fine-tuning
Open guideoMLX
Apple-silicon serving with continuous batching, memory guards, and supported SSD-backed paths
Open guidevLLM
Linux GPU servers, concurrency and OpenAI-compatible production APIs
Open guideTensorRT-LLM
Maximum NVIDIA throughput after engine tuning
Open guideSGLang
Multi-request GPU serving with prefix caching and speculative decoding
Open guideLocalAI
One OpenAI-compatible endpoint for text, vision, speech and image models, GPU optional
Open guideLemonade
Ryzen AI, Radeon and Strix Halo PCs, including the XDNA2 NPU, behind one local API
Open guideOpenVINO Model Server
Serving OpenVINO IR models on Intel CPUs, Arc GPUs and NPUs with one API
Open guideKoboldCpp
One-file GGUF chat on Windows, macOS or Linux with no Python install
Open guidellamafile
Distributing a whole GGUF model as one runnable executable on Linux, macOS, Windows and BSD
Open guideKTransformers
Very large MoE models with hot experts on the GPU and cold experts in system RAM
Open guideUnsloth
Running GGUF or MLX models from a desktop app that also trains and exports them
Open guideHow the estimates work
We start with documented model sizes and quantization, reserve space for the OS or runtime, and treat context cache as extra. MoE fit uses total parameters—not active parameters.
We separate total parameters, active MoE parameters, quantized weight fit, runtime support, and measured performance. Unknown speed stays unknown.
- 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
Questions before you download 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.