Can you run Qwen3.8 27B locally?
Yes—on the right 24GB or 32GB-class computer. The actual Q4_K_M GGUF is 19GB and its estimated text-only working set at 4K context is about 20.8GB before OS reserve. A 16GB machine cannot hold it; 32GB gives useful headroom.
Straight answer for your hardware
These are capacity estimates, not measured speed results. They use the published GGUF bytes, the official 16 full-attention-layer KV layout, and a 4K context. Your OS and frontend need additional reserve.
| Hardware memory | Verdict | Best starting point |
|---|---|---|
| 16GB | No | Q4 weights alone are 19GB |
| 24GB GPU | Yes, tight | Q4_K_M, text-only, 4K context; close other GPU apps |
| 24GB unified memory | Borderline | OS reserve makes 32GB the safer Mac/PC choice |
| 32GB | Yes | Q4_K_M with practical short/medium context headroom |
| 48GB+ | Yes | Q8_0 or Q4 with substantially longer context |
| Phone, even 24GB | No | Mobile OS/app limits leave less than the Q4 working set |
Use a current llama.cpp build. Start text-only at 4K context so you can verify the load before enabling longer context or vision. The repository's separate mmproj file is required for image/video input.
llama-server -hf ggml-org/Qwen3.8-27B-GGUF:Q4_K_M --jinja --ctx-size 409627B versus 2.4T-A95B: do not mix them up
Qwen3.8 is now a family. The 27B dense checkpoint is the practical local release. The earlier 2.4T-A95B flagship is still a server model: MoE sparsity reduces compute per token, not the amount of expert weights stored.
| Model | Smallest tracked GGUF | Realistic local tier | Best fit |
|---|---|---|---|
| Qwen3.8 27B | 19GB Q4_K_M | 24GB GPU / 32GB unified memory | Local coding, agents, vision and research |
| Qwen3.8 2.4T-A95B | 397.3GB Q1_0 | 512GB+ combined memory | Server or multi-accelerator inference |
Quick answers
Can I run Qwen3.8 27B locally?
Yes. The current Q4_K_M GGUF is 19GB and needs about 20.8GB at 4K context before operating-system reserve. A 24GB GPU can run the text model with tight settings; 32GB unified memory or VRAM is the more comfortable starting point.
Can Qwen3.8 27B run on 16GB RAM or VRAM?
Not as a complete Q4_K_M load. The weight file alone is 19GB, before KV cache, runtime buffers, the operating system, and the optional vision projector. Use a smaller model or split CPU/GPU memory with enough combined capacity.
Can Qwen3.8 27B run on a 24GB GPU?
Yes for text inference at Q4_K_M with a short context, but headroom is limited. Keep other GPU workloads closed and start at 4K context. Vision adds a separate projector and more runtime memory.
Does Qwen3.8 27B support images and video?
Yes. The official checkpoint is a native vision-language model. GGUF vision use also needs the separate mmproj file; AICanRun excludes mmproj and MTP helper files from the reported language-model weight size.
Is Qwen3.8 27B open source?
Its model weights were published by Qwen under Apache 2.0. The official card describes a 27B dense model with 262,144-token native context and optional extension to one million tokens.
What is the difference between Qwen3.8 27B and 2.4T-A95B?
The 27B model is a dense, genuinely local option for 24–48GB computers. The 2.4T-A95B model is a sparse MoE flagship whose smallest tracked GGUF is 397.3GB, so it remains server-class despite activating 95B parameters per token.
Primary source: Qwen official model card. Quant source: llama.cpp GGUF repository. Auxiliary mmproj and MTP files are excluded from weight totals. No speed number on this page is presented as an AICanRun measurement.