Can You Run Qwen3.8 Locally?

Qwen3.8-27B landed on August 5, 2026 — and it's already one of the fastest-adopted open models of the year: nearly 10,000 likes and close to a million GGUF downloads in its first ten days. It's Apache 2.0, it thinks before it answers, it reads images and video natively, and — the part this site cares about — it fits consumer GPUs remarkably well. Every number below is measured from the real GGUF files.

What's new in Qwen3.8

• Hybrid attention. Only 16 of its 64 layers keep a full KV cache; the other 48 use linear attention with a constant-size state. Long context costs a fraction of what it would on a conventional 27B.
• Thinking by default. Like the recent reasoning wave, Qwen3.8 generates a <think> chain before its final answer.
• Native vision. It understands images and video out of the box (the GGUF ships a separate small mmproj projector for that).
• 262,144-token context, extensible to 1M tokens.

How much VRAM does Qwen3.8-27B need?

Measured from the real GGUF files (27.3B parameters), at a 4,096-token context:

Your GPU	Best quant that fits in VRAM	Memory
12 GB (RTX 3060, 4070)	Q2_K_XL	≈ 11.0 GB
16 GB (4060 Ti, 5070 Ti)	IQ4_XS	≈ 15.7 GB
24 GB (3090, 4090)	Q6_K	≈ 22.4 GB
32 GB (RTX 5090)	Q8_K_XL	≈ 30.4 GB

The popular balanced pick, Q4_K_M, needs about 17 GB — comfortable on any 24 GB card with room to spare, and the reason the 16 GB tier at IQ4_XS is such good value here.

The hybrid-attention payoff: cheap long context

Because only 16 layers keep a full KV cache (with 4 KV heads at head size 256), context memory stays low:

Context	KV cache
4,096	0.25 GB
32,768	2.0 GB
131,072	8.0 GB
262,144 (full)	16.0 GB

In practice: a 24 GB card runs Q4_K_M with a 32K context at ~19 GB, and a 131K context still fits at ~25 GB with light offload — long-document work on a single consumer GPU. A conventional 27B would need several times this much for the same context.

Thinking mode and memory

Qwen3.8 thinks by default, which means it generates a long internal chain before answering. That doesn't change the model's size, but it fills your context faster — another reason its cheap KV cache matters. If you mostly want quick answers, the non-thinking mode avoids the extra tokens.

The rest of the family

Qwen3.8 also ships a flagship Qwen3.8-2.4T-A95B — 2.4 trillion parameters with 95B active. Like Kimi K3 before it, that one is data-center hardware only; the 27B is the member of the family built for machines like yours.

Qwen 3.6 vs 3.8: should you switch?

If you're running Qwen 3.6 27B today, the memory footprint is nearly identical (both fit 16 GB at ~4-bit, 24 GB at Q6). What 3.8 adds is the hybrid attention — dramatically cheaper long context — plus default thinking and native vision. For long-document or agentic use, the upgrade is worth it; for short chat, both run great.

Check it against your own hardware

See live figures for every quantization on the Qwen3.8-27B model page, adjust the context to how you actually work, or pick your graphics card. New to quant labels? Read GGUF quantization explained.

The bottom line

Qwen3.8-27B is the rare frontier-family release aimed squarely at consumer hardware: Apache 2.0, ~17 GB at Q4_K_M, vision and thinking built in, and a hybrid design that makes its 262K context genuinely usable on a single GPU. If you have 16 GB or more of VRAM, it should be on your shortlist. Check exactly what your PC can run.