你的电脑，能跑什么模型？
选设备、看适配、查部署。先找到可行的组合，再决定下载或升级。
选择硬件后，这里会显示模型适配结果。
指南与方法
完整目录与估算依据，按需展开。
全部硬件指南21
每一类设备都有不同的内存预算、推理路径和诚实上限。
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.
打开指南Mac 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.
打开指南MacBook 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.
打开指南Framework 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.
打开指南Radeon 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.
打开指南NVIDIA 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.
打开指南Mac 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.
打开指南Mac 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.
打开指南GeForce 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.
打开指南GeForce 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.
打开指南8 GB CPU laptop
Small text models with short context. Expect patient, private inference rather than speed.
打开指南16 GB CPU desktop
Comfortable with 3B–8B Q4 models; 12B is possible only with reduced context and patience.
打开指南Apple silicon · 16 GB
A polished 4B–8B local experience when the OS and apps have enough headroom.
打开指南NVIDIA · 8 GB
The mainstream 4B–8B tier. Some 12B INT4 builds fit tightly with modest context.
打开指南NVIDIA · 12 GB
Strong 8B–14B Q4 territory for a single user.
打开指南NVIDIA · 16 GB
14B models are comfortable; 20B-class low-bit MoE models are a tight upper edge.
打开指南Apple silicon · 32 GB
24B–32B Q4 models become practical for one local user.
打开指南NVIDIA · 24 GB
The consumer sweet spot for 27B and 30B-A3B INT4/Q4 models.
打开指南Apple silicon · 64 GB
32B models with long context or 70B Q4 with conservative headroom.
打开指南Workstation · 48 GB
70B Q4 serving begins here, with limited room for cache and concurrency.
打开指南Datacenter · 80 GB
Large MoE models and high-throughput servers, still bounded by quantization and cache.
打开指南全部模型指南24
Qwen3.8 Flash Next
English- 估算内存
- ~108 GB
- 上下文
- 262,144 native; 1M extended
Ling-3.0-flash
English- 估算内存
- ~74.4 GB
- 上下文
- Check selected runtime and checkpoint
Nex-N2.5-mini
English- 估算内存
- ~21 GB
- 上下文
- 262,144 (config max_position_embeddings)
Qwen3.6 27B
- 估算内存
- ~18.5 GB
- 上下文
- 262K native
Qwen3.8 27B
- 估算内存
- ~19.5 GB
- 上下文
- 262K native
Qwen3.8 Flash Next REAP-288
- 估算内存
- ~40.57 GB
- 上下文
- 7.6K recommended on measured 48 GB Mac
Qwen3.6 35B-A3B
- 估算内存
- ~25 GB
- 上下文
- 262K native
DeepSeek V4 Flash
- 估算内存
- ~176 GB
- 上下文
- 1M native
Gemma 3 1B
- 估算内存
- ~1.4 GB
- 上下文
- 32K
Llama 3.2 3B
- 估算内存
- ~2.8 GB
- 上下文
- 128K
Qwen3 4B
- 估算内存
- ~3.6 GB
- 上下文
- 32K+
Gemma 3 4B
- 估算内存
- ~4.2 GB
- 上下文
- 128K
Qwen3 8B
- 估算内存
- ~6.8 GB
- 上下文
- 32K+
Gemma 3 12B
- 估算内存
- ~9.4 GB
- 上下文
- 128K
Qwen3 14B
- 估算内存
- ~11.2 GB
- 上下文
- 32K+
gpt-oss-20b
- 估算内存
- ~16 GB
- 上下文
- 128K
Devstral Small 2 24B
- 估算内存
- ~18.5 GB
- 上下文
- 256K
Gemma 3 27B
- 估算内存
- ~20.5 GB
- 上下文
- 128K
Qwen3 30B-A3B
- 估算内存
- ~22.5 GB
- 上下文
- 32K+
Qwen3 32B
- 估算内存
- ~24.5 GB
- 上下文
- 32K+
Llama 3.1 70B
- 估算内存
- ~46 GB
- 上下文
- 128K
Llama 4 Scout
- 估算内存
- ~78 GB
- 上下文
- 10M advertised
gpt-oss-120b
- 估算内存
- ~80 GB
- 上下文
- 128K
Qwen-Image 2.1
English- 估算内存
- ~33.1 GB
- 上下文
- Not a token context: one prompt plus up to 10 reference images
为任务选择推理层15
Ollama
One-command local chat and app integration
打开指南llama.cpp
GGUF models, CPU/GPU offload, embedded and unusual hardware
打开指南LM Studio
Discovering, downloading and testing models without a terminal
打开指南MLX LM
Native Apple silicon inference, experimentation and fine-tuning
打开指南oMLX
Apple-silicon serving with continuous batching, memory guards, and supported SSD-backed paths
打开指南vLLM
Linux GPU servers, concurrency and OpenAI-compatible production APIs
打开指南TensorRT-LLM
Maximum NVIDIA throughput after engine tuning
打开指南SGLang
Multi-request GPU serving with prefix caching and speculative decoding
打开指南LocalAI
One OpenAI-compatible endpoint for text, vision, speech and image models, GPU optional
打开指南Lemonade
Ryzen AI, Radeon and Strix Halo PCs, including the XDNA2 NPU, behind one local API
打开指南OpenVINO Model Server
Serving OpenVINO IR models on Intel CPUs, Arc GPUs and NPUs with one API
打开指南KoboldCpp
One-file GGUF chat on Windows, macOS or Linux with no Python install
打开指南llamafile
Distributing a whole GGUF model as one runnable executable on Linux, macOS, Windows and BSD
打开指南KTransformers
Very large MoE models with hot experts on the GPU and cold experts in system RAM
打开指南Unsloth
Running GGUF or MLX models from a desktop app that also trains and exports them
打开指南估算是怎么算的
从官方模型大小和量化格式出发，为系统与运行时留出空间，并把上下文缓存视为额外开销。MoE 按总参数计算，而不是激活参数。
我们区分总参数、MoE 激活参数、量化权重、运行时支持和实测速率；没有证据的速度保持未知。
- 内存适配是估算，不是速度测试。
- MoE 按总参数计算；共享内存与分布式内存都需要额外核对。
- 厂商口径、社区实测、运行条件与未知项分别披露。
目录数据更新时间: 2026-10-03
下载 40GB 之前先问清楚
内存够用，就一定能运行吗？
不一定。容量只是一项估算，还要核对量化文件、运行时、操作系统、驱动和上下文缓存。共享内存也要为系统与其他程序留出空间。
MoE 的激活参数就是内存需求吗？
不是。激活参数影响每个 token 的计算量；存储或卸载权重仍需按总参数和实际量化文件计算。
选多个模型，代表可以同时运行吗？
不代表。这里比较的是每个模型分别运行时的适配情况；同时驻留、长上下文与多请求都需要额外内存和独立测试。
为什么有些速度显示未知？
只有可追溯、注明设备、模型版本和运行条件的报告才展示数字。预填充速度、单流生成速度和多流总吞吐分别列出，不能互相替代。