Text Gen · Model Guide · 5 min read

Qwen3.8 27B

Apache-2.0 dense model with hybrid attention — one of the best quality-per-VRAM buys on a 16–24 GB GPU.

Basic Information

Qwen3.8 27B is Alibaba's mid-size dense model in the Qwen3.8 generation, and a favorite of the local-AI community for a simple reason: it delivers near-frontier instruction-following, coding, and multilingual quality at a size that a single 16–24 GB GPU can genuinely host. It is fully Apache 2.0 licensed, which makes it the default answer for commercial local deployments.

Architecturally it is interesting: only every fourth layer keeps a full KV-cache (hybrid attention), which slashes long-context memory cost — the model supports 262K-token contexts without the KV-cache ballooning the way older designs would.

Best use cases

• Commercial local assistants and internal tools (Apache 2.0 removes licensing friction)
• Coding help and technical writing on a single workstation GPU
• Long-document work — 262K context with unusually low cache cost
• Fine-tuning experiments: small enough to train (LoRA) on one card

Computational Requirements

Comfortably local: Q4 lands around 13 GB of weights plus a very cheap KV-cache, fitting well inside a 24 GB card with generous context.

• 16 GB cards: use Q4 and keep context moderate. 24 GB cards: Q4 with large context, or Q8 at shorter context.

Commercial license

Apache 2.0 — the most business-friendly license in AI: commercial use, modification, redistribution, and even closed derivatives are all permitted with attribution. If licensing anxiety is the thing stopping your local deployment, this model is the answer.