С чего вы начинаете?
Начните с имеющегося оборудования или нужных моделей. Сравните вместимость, скорость и доказательства до развёртывания.
Select hardware to see model fit results.
Guides & methodology
Browse the complete catalogs and estimation method when you need them.
Всё оборудование21
У каждого класса свой бюджет памяти, путь запуска и честный предел.
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.
Открыть гидMac 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.
Открыть гидMacBook 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.
Открыть гидFramework 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.
Открыть гидRadeon 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.
Открыть гидNVIDIA 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.
Открыть гидMac 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.
Открыть гидMac 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.
Открыть гидGeForce 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.
Открыть гидGeForce 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.
Открыть гид8 GB CPU laptop
Small text models with short context. Expect patient, private inference rather than speed.
Открыть гид16 GB CPU desktop
Comfortable with 3B–8B Q4 models; 12B is possible only with reduced context and patience.
Открыть гидApple silicon · 16 GB
A polished 4B–8B local experience when the OS and apps have enough headroom.
Открыть гидNVIDIA · 8 GB
The mainstream 4B–8B tier. Some 12B INT4 builds fit tightly with modest context.
Открыть гидNVIDIA · 12 GB
Strong 8B–14B Q4 territory for a single user.
Открыть гидNVIDIA · 16 GB
14B models are comfortable; 20B-class low-bit MoE models are a tight upper edge.
Открыть гидApple silicon · 32 GB
24B–32B Q4 models become practical for one local user.
Открыть гидNVIDIA · 24 GB
The consumer sweet spot for 27B and 30B-A3B INT4/Q4 models.
Открыть гидApple silicon · 64 GB
32B models with long context or 70B Q4 with conservative headroom.
Открыть гидWorkstation · 48 GB
70B Q4 serving begins here, with limited room for cache and concurrency.
Открыть гидDatacenter · 80 GB
Large MoE models and high-throughput servers, still bounded by quantization and cache.
Открыть гидВсе модели24
Qwen3.8 Flash Next
English- Оценка памяти
- ~108 GB
- Контекст
- 262,144 native; 1M extended
Ling-3.0-flash
English- Оценка памяти
- ~74.4 GB
- Контекст
- Check selected runtime and checkpoint
Nex-N2.5-mini
English- Оценка памяти
- ~21 GB
- Контекст
- 262,144 (config max_position_embeddings)
Qwen3.6 27B
- Оценка памяти
- ~18.5 GB
- Контекст
- 262K native
Qwen3.8 27B
- Оценка памяти
- ~19.5 GB
- Контекст
- 262K native
Qwen3.8 Flash Next REAP-288
- Оценка памяти
- ~40.57 GB
- Контекст
- 7.6K recommended on measured 48 GB Mac
Qwen3.6 35B-A3B
- Оценка памяти
- ~25 GB
- Контекст
- 262K native
DeepSeek V4 Flash
- Оценка памяти
- ~176 GB
- Контекст
- 1M native
Gemma 3 1B
- Оценка памяти
- ~1.4 GB
- Контекст
- 32K
Llama 3.2 3B
- Оценка памяти
- ~2.8 GB
- Контекст
- 128K
Qwen3 4B
- Оценка памяти
- ~3.6 GB
- Контекст
- 32K+
Gemma 3 4B
- Оценка памяти
- ~4.2 GB
- Контекст
- 128K
Qwen3 8B
- Оценка памяти
- ~6.8 GB
- Контекст
- 32K+
Gemma 3 12B
- Оценка памяти
- ~9.4 GB
- Контекст
- 128K
Qwen3 14B
- Оценка памяти
- ~11.2 GB
- Контекст
- 32K+
gpt-oss-20b
- Оценка памяти
- ~16 GB
- Контекст
- 128K
Devstral Small 2 24B
- Оценка памяти
- ~18.5 GB
- Контекст
- 256K
Gemma 3 27B
- Оценка памяти
- ~20.5 GB
- Контекст
- 128K
Qwen3 30B-A3B
- Оценка памяти
- ~22.5 GB
- Контекст
- 32K+
Qwen3 32B
- Оценка памяти
- ~24.5 GB
- Контекст
- 32K+
Llama 3.1 70B
- Оценка памяти
- ~46 GB
- Контекст
- 128K
Llama 4 Scout
- Оценка памяти
- ~78 GB
- Контекст
- 10M advertised
gpt-oss-120b
- Оценка памяти
- ~80 GB
- Контекст
- 128K
Qwen-Image 2.1
English- Оценка памяти
- ~33.1 GB
- Контекст
- Not a token context: one prompt plus up to 10 reference images
Выберите слой инференса15
Ollama
One-command local chat and app integration
Открыть гидllama.cpp
GGUF models, CPU/GPU offload, embedded and unusual hardware
Открыть гидLM Studio
Discovering, downloading and testing models without a terminal
Открыть гидMLX LM
Native Apple silicon inference, experimentation and fine-tuning
Открыть гидoMLX
Apple-silicon serving with continuous batching, memory guards, and supported SSD-backed paths
Открыть гидvLLM
Linux GPU servers, concurrency and OpenAI-compatible production APIs
Открыть гидTensorRT-LLM
Maximum NVIDIA throughput after engine tuning
Открыть гидSGLang
Multi-request GPU serving with prefix caching and speculative decoding
Открыть гидLocalAI
One OpenAI-compatible endpoint for text, vision, speech and image models, GPU optional
Открыть гидLemonade
Ryzen AI, Radeon and Strix Halo PCs, including the XDNA2 NPU, behind one local API
Открыть гидOpenVINO Model Server
Serving OpenVINO IR models on Intel CPUs, Arc GPUs and NPUs with one API
Открыть гидKoboldCpp
One-file GGUF chat on Windows, macOS or Linux with no Python install
Открыть гидllamafile
Distributing a whole GGUF model as one runnable executable on Linux, macOS, Windows and BSD
Открыть гидKTransformers
Very large MoE models with hot experts on the GPU and cold experts in system RAM
Открыть гидUnsloth
Running GGUF or MLX models from a desktop app that also trains and exports them
Открыть гидКак считается оценка
Берём документированные размеры и квантование, оставляем запас системе, а кэш считаем отдельно. Для MoE используются все параметры.
Мы разделяем общее число параметров, активные MoE, размер квантования, поддержку и измеренную скорость.
- 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
До загрузки 40 ГБ
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.