Which large model fits your laptop?

Check the RAM, disk and operating-system paths for native C Qwen3.8-Flash-Next, Qwen3.8-27B and DeepSeek-V4-Flash-0731 CPU inference.

Windows environment

Qwen3.8-Flash-Next

Checking
  • Minimum path: 8 GB RAM, about 75 GB free disk; selected automatically
  • Model: best under 200B, 125B-A6B main model + 51B PLE
  • Reference speed: 5.03 token/s resident chat; 9.89 positions/s exact batch
  • Use it: terminal chat or resident OpenAI-compatible API with function tools
  • Runtime: native C, direct split GGUF, no GPU or Python
Open Flash-Next Quick Start

Qwen3.8-27B

Checking
  • Minimum tested: 8 GB RAM, about 10 GB free disk
  • Higher-quality path: 20 GB available memory, about 17 GB disk
  • Reference speed: up to 2.52 token/s on a 32 GB x86 laptop
  • Use it: terminal chat or resident OpenAI-compatible API with function tools
  • Runtime: native C, direct GGUF, no GPU or Python
Open Qwen Quick Start

DeepSeek-V4-Flash-0731

Checking
  • Minimum tested: 8 GB RAM, about 172 GB free disk
  • Checkpoint: native 284B-A13B, streamed from SSD
  • Reference speed: ordinary 0.59 token/s; best prompt lookup 1.12 token/s
  • Use it: terminal chat or resident OpenAI-compatible API with function tools
  • Runtime: pure C, no GPU, Python or weight conversion
Open DeepSeek Quick Start

Real output from the published engines

Qwen3.8-27B real sample output
Qwen3.8-27B, native C, laptop CPU.
DeepSeek-V4-Flash-0731 real sample output
DeepSeek-V4-Flash-0731, pure C, laptop CPU.

Questions before downloading / 下载前常见问题

Which models can an 8 GB laptop run? / 8 GB 笔记本能运行哪些模型？

All three engines have an 8 GB path. Qwen3.8-Flash-Next automatically uses bounded-memory weight streaming; 12 GB or more uses its faster default path.

Does the 8 GB path preserve original model precision? / 低内存方案会损失精度吗？

Qwen IQ1_M is a lossy quantization of the original checkpoint, while its runtime speedups add no further approximation for the same GGUF. DeepSeek reads the released native FP8/FP4 checkpoint without weight conversion.

Why does DeepSeek need 172 GB disk but only 8 GB RAM? / 为什么磁盘很大、内存却可以很小？

The complete 284B-A13B checkpoint stays on disk. The engine keeps a bounded RAM cache and streams cold MoE expert weights from SSD only when routing needs them.

Can I run this on Windows? / Windows 可以运行吗？

Yes, through WSL2. Linux and macOS provide the required POSIX interfaces directly.